Allanus Tsoi, David Nualart, George Yin Eds. Stochastic Analysis, Stochastic Systems, and Applications To Finance PDF

Stochastic Analysis,
Stochastic Systems, and

Applications to Finance
8197.9789814355704-tp.indd 1 5/19/11 12:05 PM

Stochastic Analysis,
Stochastic Systems, and
Applications to Finance
Edited by
Allanus Tsoi
University of Missouri, Columbia, USA
David Nualart
University of Kansas, USA
George Yin
Wayne State University, Michigan, USA
World Scientific
NEW JERSEY • LONDON • SINGAPORE • BEIJING • SHANGHAI • HONG KONG • TA I P E I • CHENNAI
8197.9789814355704-tp.indd 2 5/19/11 12:05 PM

This page is intentionally left blank
Published by
World Scientific Publishing Co. Pte. Ltd.
5 Toh Tuck Link, Singapore 596224
USA office: 27 Warren Street, Suite 401-402, Hackensack, NJ 07601
UK office: 57 Shelton Street, Covent Garden, London WC2H 9HE
British Library Cataloguing-in-Publication Data

A catalogue record for this book is available from the British Library.
STOCHASTIC ANALYSIS, STOCHASTIC SYSTEMS, AND APPLICATIONS

TO FINANCE
Copyright © 2011 by World Scientific Publishing Co. Pte. Ltd.
All rights reserved. This book, or parts thereof, may not be reproduced in any form or by any means,
electronic or mechanical, including photocopying, recording or any information storage and retrieval
system now known or to be invented, without written permission from the Publisher.
For photocopying of material in this volume, please pay a copying fee through the Copyright
Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to
photocopy is not required from the publisher.
ISBN-13 978-981-4355-70-4
ISBN-10 981-4355-70-4
Printed in Singapore.
He Yue - Stochastic Analysis.pmd 1 5/11/2011, 3:59 PM

May 13, 2011 11:8 WSPC - Proceedings Trim Size: 9in x 6in cnts
Contents
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
Contributors and Addresses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
Part I. Stochastic Analysis and Systems
1. Multidimensional Wick-Itô Formula for Gaussian Processes . . . . . . . . 3

D. Nualart and S. Ortiz-Latorre
2. Fractional White Noise Multiplication . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
A. H. Tsoi
3. Invariance Principle of Regime-Switching Diffusions . . . . . . . . . . . . . . . 43
C. Zhu and G. Yin
Part II. Finance and Stochastics
4. Real Options and Competition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

A. Bensoussan, J. D. Diltz, and S. R. Hoe
5. Finding Expectations of Monotone Functions of Binary Random
Variables by Simulation, with Applications to Reliability,
Finance, and Round Robin Tournaments . . . . . . . . . . . . . . . . . . . . . . . . . 101
M. Brown, E. A. Peköz, and S. M. Ross
6. Filtering with Counting Process Observations and Other
Factors: Applications to Bond Price Tick Data . . . . . . . . . . . . . . . . . . . 115
X. Hu, D. R. Kuipers, and Y. Zeng
May 13, 2011 11:8 WSPC - Proceedings Trim Size: 9in x 6in cnts
vi Contents
7. Jump Bond Markets Some Steps towards General Models

in Applications to Hedging and Utility Problems . . . . . . . . . . . . . . . . . 145
M. Kohlmann and D. Xiong
8. Recombining Tree for Regime-Switching Model: Algorithm
and Weak Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .193
R. H. Liu
9. Optimal Reinsurance under a Jump Diffusion Model . . . . . . . . . . . . . 215
S. Luo
10. Applications of Counting Processes and Martingales in
Survival Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
J. Sun
11. Stochastic Algorithms and Numerics for Mean-Reverting
Asset Trading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
Q. Zhang, C. Zhuang, and G. Yin
April 27, 2011 9:46 WSPC - Proceedings Trim Size: 9in x 6in 01-preface
vii
Preface
This volume contains 11 chapters. It is an expanded version of the papers

presented at the first Kansas–Missouri Winter School of Applied Probabil-
ity, which was organized by Allanus Tsoi and was held at the University of
Missouri, February 14 and 15, 2008. It brought together researchers from
different parts of the country to review and to update the recent advances,
and to identify future directions in the areas of applied probability, stochas-
tic processes, and their applications.
After the successful conference was over, there was a strong support
of publishing the papers delivered in the conference as an archival volume.
Based on the support, we began the preparation on this project. In addition
to papers reported at the conference, we have invited a number of colleagues
to contribute additional papers.
As an archive, this volume presents some of the highlights of the con-
ference, as well as some of most recent developments in stochastic systems
and applications. This book is naturally divided into two parts. The first
part contains some recent results in stochastic analysis, stochastic processes
and related fields. It explores the Itô formula for multidimensional Gaussian
processes using the Wick integral, introduces the notion of fractional white
noise multiplication, and discusses the LaSalle type of invariance principles
for hybrid switching diffusions. The second part of the book is devoted to fi-
nancial mathematics, insurance models, and applications. Included here are
optimal investment policies for irreversible capital investment projects un-
der uncertainty in monopoly and Stackelberg leader-follower environments,
April 27, 2011 9:46 WSPC - Proceedings Trim Size: 9in x 6in 01-preface
viii Preface
finding expectations of monotone functions of binary random variables by

simulation, with applications to reliability, finance, and round robin tour-
naments, jump bond markets with general models in applications to hedg-
ing and utility problems, algorithm and weak convergence for recombining
tree in a regime-switching model, applications of counting processes and
martingales in survival analysis, extended filtering micro-movement model
with counting process observations and applications to bond price tick data,
optimal reinsurance for a jump diffusion model, recursive algorithms and
numerical studies for mean-reverting asset trading.
Without the encouragement and assistance of many colleagues, this vol-
ume would have never come into being. We thank all the authors of this
volume, and all of the speakers of the conference for their contributions. The
financial support provided by the University of Missouri for this conference
is also greatly acknowledged.
Allanus Tsoi
Columbia, Missouri
David Nualart
Lawrence, Kansas
George Yin
Detroit, Michigan
April 21, 2011 16:30 WSPC - Proceedings Trim Size: 9in x 6in names
ix
Contributors and Addresses
• Alain Bensoussan, School of Management, University of Texas at

Dallas, Richardson, TX 75083-0688, USA. & The Hong Kong Poly-
technic University, Hong Kong. Email: alain.bensoussan@utdallas.
edu
• Mark Brown, Department of Mathematics, City College, CUNY,
New York, NY, USA. Email: cybergarf@aol.com
• J. David Diltz, Department of Finance and Real Estate, The Uni-
versity of Texas at Arlington, Arlington, TX 76019, USA. Email:
diltz@uta.edu
• SingRu Hoe, School of Management, University of Texas at Dallas,
Richardson, TX 75083-0688, USA. Email: celinehoe@utdallas.edu
• Xing Hu, Department of Economics, Princeton University, Prince-
ton, 08544, USA. Email: xinghu@princeton.edu
• Michael Kohlmann, Department of Mathematics and Statistics,
University of Konstanz, D-78457, Konstanz, Germany. Email:
michael.kohlmann@uni-konstanz.de
• David R. Kuipers, Department of Finance, Henry W. Bloch
School of Business and Public Administration, University of Mis-
souri at Kansas City, Kansas City, MO 64110, USA. Email:
kuipersd@umkc.edu
• Ruihua Liu, Department of Mathematics, University of Dayton,
300 College Park, Dayton, OH 45469-2316, USA. Email: rui-
hua.liu@notes.udayton.edu
April 21, 2011 16:30 WSPC - Proceedings Trim Size: 9in x 6in names
x Contributors and Addresses
• Shangzhen Luo, Department of Mathematics, University of North-

ern Iowa, Cedar Falls, Iowa, 50614-0506, USA. Email: luos@uni.edu
• David Nualart, Department of Mathematics, University of Kansas,
Lawrence, KS 66045, USA. Email: nualart@math.ku.edu
• Salvador Ortiz-Latorre, Departament de Probabilitat, Lògica i Es-
tadı́stica, Universitat de Barcelona, Gran Via 585, 08007 Barcelona,
Spain. Email: sortiz@ub.edu
• Erol A. Pekoz, School of Management, Boston University,
595 Commonwealth Avenue, Boston, MA 02215, USA. Email:
pekoz@bu.edu
• Sheldon M. Ross, Department of Industrial and Systems Engineer-
ing, University of Southern California, Los Angeles, CA 90089,
USA. Email: smross@usc.edu
• Jianguo Sun, Department of Statistics, University of Missouri,
USA. Email: sunj@missouri.edu
• Allanus Hak-Man Tsoi, Department of Mathematics, University of
Missouri, Columbia, MO 65211, USA. Email: tsoia@missouri.edu
• Dewen Xiong, Department of Mathematics, Shanghai Jiaotong
University, Shanghai 200240, People’s Republic of China. Email:
xiongdewen@sjtu.edu.cn
• George Yin, Department of Mathematics, Wayne State University,
Detroit, MI 48202, USA. Email: gyin@math.wayne.edu
• Yong Zeng, Department of Mathematics and Statistics, University
of Missouri at Kansas City, Kansas City, MO 64110, USA. Email:
zengy@umkc.edu
• Qing Zhang, Department of Mathematics, University of Georgia,
Athens, GA 30602, USA. Email: qingz@math.uga.edu
• Chao Zhu, Department of Mathematical Sciences, University
of Wisconsin-Milwaukee, Milwaukee, WI 53201, USA. Email:
zhu@uwm.edu
• Chao Zhuang, Marshall School of Business, University of Southern
California, Los Angeles, CA 90089, USA. Email: czhuang@usc.edu
May 11, 2011 10:49 WSPC - Proceedings Trim Size: 9in x 6in 01-nu
Multidimensional Wick-Itô Formula for Gaussian Processes
D. Nualart∗
Department of Mathematics, University of Kansas
Lawrence, KS 66045, USA
E-mail: nualart@math.ku.edu
S. Ortiz-Latorre
Departament de Probabilitat, Lògica i Estadı́stica, Universitat de Barcelona
Gran Via 585, 08007 Barcelona, Spain
Email: sortiz@ub.edu
An Itô formula for multidimensional Gaussian processes using the Wick inte-
gral is obtained. The conditions allow us to consider processes with infinite
quadratic variation. As an example we consider a correlated heterogenous frac-
tional Brownian motion. We also use this Itô formula to compute the price of
an exchange option in a Wick-fractional Black-Scholes model.
Keywords: Wick-Itô formula; Gaussian processes; Malliavin calculus.
1. Introduction
The classical stochastic calculus and Itô’s formula can be extended to semi-
martingales. There has been a recent interest in developing a stochastic
calculus for Gaussian processes which are not semimartingales such as the
fractional Brownian motion (fBm for short). These developments are mo-
tivated by the fact that fBm and other related processes are suitable input
noises in practical problems arising in a variety of fields including finance,
telecommunications and hydrology (see, for instance, Mandelbrot and Van
Ness7 and Sottinen13 ).
A possible definition of the stochastic integral with respect to the fBm
is based on the divergence operator appearing in the stochastic calculus
of variations. This approach to define stochastic integrals started from the
work by Decreusefond and Üstünel3 and was further developed by Carmona
∗ Supported by the NSF Grant DMS-0604207

4 D. Nualart and S. Ortiz-Latorre
and Coutin2 and Duncan, Hu and Pasik-Duncan4 (see also Hu5 and Nu-
alart9 for a general survey papers on the stochastic calculus for the fBm).
The divergence integral can be approximated by Riemman sums defined
using the Wick product, and it has the important property of having zero
expectation.
Nualart and Taqqu11,12 have proved a Wick-Itô formula for general
Gaussian processes. In 11 they have considered Gaussian processes with
finite quadratic variation, which includes the fBm with Hurst parameter
H > 1/2. The paper 12 deals with the change-of-variable formula for Gaus-
sian processes with infinite quadratic variation, in particular the fBm with
Hurst parameter H ∈ (1/4, 1/2). The lower bound for H is a natural one,
see Alòs, Mazet and Nualart.1
The aim of this paper is to generalize the results of Nualart and Taqqu12
to the multidimensional case. We introduce the multidimensional Wick-Itô
integral as a limit of forward Riemann sums and prove a Wick-Itô formula
under conditions similar to those in Nualart and Taqqu,12 allowing infinite
quadratic variation processes.
The paper is organized as follows. In Section 2, we introduce the con-
ditions that the multidimensional Gaussian process must satify and state
the Itô formula. Section 3 contains some definitions in order to introduce
the Wick integral. In Section 4 we prove some technical lemmas using ex-
tensively the integration by parts formula for the derivative operator. The
convergence results used in the proof of the main theorem are proved in
Section 5. Section 6 is devoted to study two examples related to the multi-
dimensional fBm with parameter H > 1/4. Finally, in section 7 we use our
Itô formula to compute the price of an exchange option on a Wick-fractional
Black-Scholes market.
2. Preliminaries
Let X = {Xt , t ∈ [0, T ]} be a d-dimensional centered Gaussian process with
continuous covariance function matrix R(s, t), that is,
Ri,j (s, t) = E[Xsi Xtj ],
for i, j = 1, . . . , d. For s = t, we have the covariance matrix Vt = R(t, t).

We denote by H be the space obtained as the completion of the set of
step functions in A = [0, T ] × {1, . . . , d} with respect the scalar product
h1i[0,s] , 1j[0,t] iH = Ri,j (s, t) , 0 ≤ s, t ≤ T, 1 ≤ i, j ≤ d,

Multidimensional Wick-Itô Formula for Gaussian Processes 5
where
1i[0,s] = 1[0,s]×{i} (x, k) , (x, k) ∈ A.
The mapping 1i[0,t] 7→ Xti can be extended to a linear isometry between the
space H and the Gaussian Hilbert space generated by the process X. We
denote by let I1 : h 7→ X (h) , h ∈ H this isometry.
Let H⊗m denote the mth tensor power of H, equipped with the following
scalar product
m
Y
hh1 ⊗ · · · ⊗ hm , g1 ⊗ · · · ⊗ gm iH⊗m = hhi , gi iH ,
i=1
where h1 , . . . , hm , g1 , . . . , gm ∈ H. The subspace of mth symmetric tensors

will be denoted by Hm . In Hm we introduce the modified scalar prod-
uct given by h·, ·iHm = m! h·, ·iH⊗m . In this way, the multiple stochastic
integral Im is an isometry between Hm and the mth Wiener chaos (see
Nualart10 and also Janson6 for a more detailed discussion of tensor prod-
ucts of Hilbert spaces). We denote by h1 · · · hm the symmetrization of
the tensor product h1 ⊗ · · · ⊗ hm .
Now consider the set of smooth random variables S. A random variable
F ∈ S has the form
F = f (X (h1 ) , . . . , X (hn )) , (1)
with h1 , . . . , hn ∈ H, n ≥ 1, and f ∈ Cb∞ (Rn ) (f and all its partial deriva-

tives are bounded). In S one can define the derivative operator D as
n
X
DF = ∂i f (X (h1 ) , . . . , X (hn )) hi ,
i=1
which is an element of L2 (Ω; H). By iteration one obtains

n
X ∂mf
Dm F = (X (h1 ) , . . . , X (hn )) hi1 ⊗ · · · ⊗ him ,
i1 ,...,im =1
∂xi1 · · · ∂xim
which is an element of L2 (Ω; Hm ).
Definition 2.1. For m ≥ 1, the space Dm,2 is the completion of S with

respect to the norm kF km,2 defined by
m
2
kF k2m,2 = E[F 2 ] +
X
E[ Di F H⊗i ].

i=1
The Wick product F X (h) between a random variable F ∈ D1,2 and

the Gaussian random variable X (h) is defined as follows.
Definition 2.2. Let F ∈ D1,2 and h ∈ H. Then the Wick product F X (h)
is defined by
F X (h) = F X (h) − hDF, hiH .
Actually, the Wick product coincides with the divergence (or the Sko-
rohod integral) of F h, and by the properties of the divergence operator we
can write
E [F X (h)] = E [hDF, hiH ] . (2)
The Wick integral of a stochastic process u with respect to X is defined

as the limit of Riemann sums constructed using the Wick product. For this
we need some notation. Denote by D the set of all partitions of [0, T ]
π = {0 = t0 < t1 < · · · < tn = T }
such that
|π|
≤ D,
|π|inf
where |π| = max0≤i≤n−1 (ti+1 − ti ) , |π|inf = min0≤i≤n−1 (ti+1 − ti ) , and

D is a positive constant.
Definition 2.3. Let u = {ut , t ∈ [0, T ]} be a d-dimensional stochastic pro-

cess such that uit ∈ D1,2 for all t ∈ [0, T ] and i = 1, . . . , d. The Wick
integral
Z T d Z T
ujt dXtj
X
ut dXt =
0 j=1 0
is defined as the limit in probability, if it exists, of the forward Riemann

sums
d n−1
ujti (Xtji+1 − Xtji )
X X
j=1 i=0
as |π| tends to zero, where π runs over all the partitions of the interval [0, T ]
in the class D.
3. Main Result
We will make use of the following assumptions.
Assumptions.
(A1) For all j, k ∈ {1, . . . , d} the function t 7→ Vtj,k has bounded varia-
tion on [0, T ].
(A2) For all k, l ∈ {1, . . . , d}
n−1
E[∆i X k ∆j X l ]2 → 0,
X
as |π| → 0.
i,j=0
(A3) For all j, k ∈ {1, . . . , d}

n−1
X 2
sup E[Xsj ∆i X k ] → 0,

as |π| → 0,
i=0 0≤s≤t
where ∆i X j = Xtji+1 − Xtji , and π runs over all partitions of [0, T ]

in the class D.
Our purpose is to derive a change-of-variable formula for the process

f (Xt ), where f : Rd → R if a function satisfying the following condition.
(A4) For every multi-index α = (α1 , ..., αd ) ∈ Nd with |α| := α1 + · · · +

αd ≤ 7, the iterated derivatives
∂ |α| f
∂ α f (x) = (x)
∂xα
1
1
· · · ∂xα
d
d
exist, are continuous, and satisfy

h i
sup E |∂ α f (Xt )|2 < ∞. (3)
t∈[0,T ]
Condition (3) holds if det Vt > 0 for all t ∈ (0, T ], and the partial
derivatives ∂ α f satisfy the exponential growth condition
2
|∂ α f (x)| ≤ CT ecT |x| , (4)
for all t ∈ [0, T ] , x ∈ Rd , where CT > 0 and cT are such that

1 xT Vt−1 x
0 < cT < inf 2 <∞ (5)
4 0<t≤T |x|
x∈Rd ,|x|>0
(see Lemma 4.5).

Besides the multi-index notation for the derivatives, we will also use the
following notation for iterated derivatives. Let f (x1 , . . . , xd ) be a sufficiently
smooth function, then
∂f
∂i f = , i ∈ {1, . . . , d}
∂xi
∂im1 ,...,im f = ∂im ∂im−1 (· · · ∂i2 (∂i1 f ) ,

ik ∈ {1, . . . , d} , k = 1, . . . , m.
The next theorem is the main result of the paper.
Theorem 3.1. Suppose that the Gaussian process X and the function f
satisfy the preceding assumptions (A1) to (A4). Then the forward integrals
(see Definition 2.3)
Z t
∂j f (Xs ) dXsj , 0 ≤ t ≤ T, j = 1, . . . , d
0
exist and the following Wick-Itô formula holds:

d Z t d Z
X 1 X t 2
f (Xt ) = f (X0 ) + ∂j f (Xs ) dXsj + ∂j,k f (Xs ) dVsj,k .
j=1 0
2 0 j,k=1
Proof. Using the Taylor expansion of f up to fourth order in two consec-

utive points of a partition π = {0 = t0 < t1 < · · · < tn = t} in the class D
we obtain
d d
X 1 X 2
∂j f (Xti ) ∆i X j + ∂j,k f (Xti ) ∆i X j ∆i X k

f Xti+1 = f (Xti ) +
j=1
2
j,k=1
1 1
+ T3π (i) + T4π (i) ,
3! 4!
where
d
X
T3π (i) = 3
∂j,k,l f (Xti ) ∆i X j ∆i X k ∆i X l ,
j,k,l=1
d
X
T4π (i) = 4
∂j,k,l,m f (X i )∆i X j ∆i X k ∆i X l ∆i X m ,
j,k,l,m=1
and
X i = λXti + (1 − λ) Xti+1 , 0 ≤ λ ≤ 1.
By the definition of the Wick product, see Definition 2.2, one has
∂j f (Xti ) ∆i X j = ∂j f (Xti ) ∆i X j + hD (∂j f (Xti )) , 1jδi iH ,
where δi = (ti , ti+1 ]. Taking into account that
d
X
2
D (∂j f (Xti )) = ∂j,k f (Xti ) 1k[0,ti ] ,
k=1
one gets
d d d
f (Xti ) h1k[0,ti ] , 1jδi iH .
X X X
∂j f (Xti ) ∆i X j = ∂j f (Xti ) ∆i X j + 2
∂j,k
j=1 j=1 j,k=1
Using the definition of h·, ·iH and adding and subtracting 21 E ∆i X j ∆i X k

we have
1 j,k 1
h1k[0,ti ] , 1jδi iH = E[Xtki (Xtji+1 − Xtji )] = ϕi − E ∆i X j ∆i X k ,

2 2
where
h i
ϕj,k
i =E Xtji+1 − Xtji Xtki+1 + Xtki .
This gives
d
X
∂j f (Xti ) ∆i X j

f Xti+1 = f (Xti ) +
j=1
d
1 X
2
f (Xti ) ∆i X j ∆i X k − E ∆i X j ∆i X k

+ ∂j,k
2
j,k=1
d
1 X 2
+ ∂j,k f (Xti ) ϕj,k π π
i + T3 (i) + T4 (i) .
2
j,k=1
Hence,
n−1
X
f (Xt ) = f (X0 ) + f Xti+1 − f (Xti )
i=0
n−1
XX d
= f (X0 ) + ∂j f (Xti ) ∆i X j
i=0 j=1
n−1 d
1X X 2 1 π 1 π 1
+ ∂j,k f (Xti ) ϕj,k
i + R2 + R + Rπ ,
2 i=0 2 3! 3 4! 4
j,k=1
where
n−1
X d
X
R2π = 2
f (Xti ) ∆i X j ∆i X k − E ∆i X j ∆i X k

∂j,k
i=0 j,k=1
n−1
X n−1
X d
X
R3π = T3π (i) = 3
i=0 i=0 j,k,l=1
n−1
X n−1
X d
X
R4π = T4π (i) = 4
∂j,k,l,m f (X i )∆i X j ∆i X k ∆i X l ∆i X m .
i=0 i=0 j,k,l,m=1
Note that
d d
1 X 2 1X 2
∂j,k f (Xti ) ϕj,k = ∂ f (Xti ) ϕj,j
2 i
2 j=1 j,j i
j,k=1
d
1 X 2
+ ∂j,k f (Xti ) (ϕj,k k,j
i + ϕi )
2
k>j=1
d
1X 2
= ∂ f (Xti ) (Vtj,j − Vtj,j )
2 j=1 j,j i+1 i
d
f (Xti ) (Vtj,k − Vtj,k
X
2
+ ∂j,k i+1 i
)
k>j=1
d
1 X 2
= ∂j,k f (Xti ) (Vtj,k − Vtj,k ).
2 i+1 i
j,k=1
Using Assumption (A1) it is easy to show the almost sure convergence

n−1 d d
1X X 2 1 X 2
lim ∂j,k f (Xti ) (Vtj,k − V j,k
ti ) = ∂j,k f (Xs ) dVsj,k
|π|→0 2 i+1
2
i=0 j,k=1 j,k=1
as |π| → 0. The convergences of R2π and R3π to zero in L2 (Ω) as |π| → 0 are
proved in Propositions 5.1 and 5.2. The convergence of R4π to zero in L1 (Ω)
as |π| → 0 is proved in Proposition 5.3. This clearly implies the convergence
in probability
n−1
XX d d Z
X t
j
lim ∂j f (Xti ) ∆i X = ∂j f (Xs ) dXsj ,
|π|→0 0
i=0 j=1 j=1
and the result follows.

Remark 3.1. We can also consider a function f (t, x) depending on time

such that the partial derivative ∂f (t, x) exists and is continuous. In this
∂t R
t
case we obtain the additional term 0 ∂f ∂t (s, Xs )ds.
In order to prove Propositions 5.1, 5.2 and 5.3 we need to introduce

some technical concepts and prove a number of lemmas.
4. Technical Lemmas
In this section we establish some preliminary lemmas. The first one is trivial.
Lemma 4.1. Let F ∈ Dm+n,2 and h1 , . . . , hm , g1 , . . . , gn ∈ H. Then

hDn hDm F, h1 · · · hm iH⊗m , g1 · · · gn iH⊗n
= Dm+n F, h1 · · · hm g1 · · · gn H⊗m+n .

The next lemmas are based on the integration by parts formula.
Lemma 4.2. Let F ∈ D2,2 and h, g ∈ H. Then

E [F X (h) X (g)] = E[ D2 F, h g H⊗2 ] + E [F ] hh, giH .

Proof. See Nualart and Taqqu12 , Lemma 6.
Lemma 4.3. Let F ∈ D2,2 , h, g ∈ H, ξ = X (h) X (g) − hh, giH . Then

E [F ξ] = E[ D2 F, h g H⊗2 ].

Proof. It is an immediate consequence of the preceding lemma.
Lemma 4.4. Let F ∈ D4,2 , h1 , h2 , g1 , g2 ∈ H, ξ1 = X (h1 ) X (g1 ) −

hh1 , g1 iH and ξ2 = X (h2 ) X (g2 ) − hh2 , g2 iH . Then
E [F ξ1 ξ2 ] = E[ D4 F, h2 g2 h1 g1 H⊗4 ] + E[ D2 F, h2 g1 H⊗2 ] hh1 , g2 iH

+ E[ D2 F, g1 g2 H⊗2 ] hh1 , h2 iH + E[ D2 F, h1 h2 H⊗2 ] hg1 , g2 iH

+ E[ D2 F, h1 g2 H⊗2 ] hh2 , g1 iH + 2E [F ] hh1 g1 , h2 g2 iH⊗2 .

Proof. Applying the last lemma with F replaced by F ξ1 and ξ by ξ2 , we

get
E [F ξ1 ξ2 ] = E[ D2 (F ξ1 ) , h2 g2 H⊗2 ].

Now, by the Leibniz rule for the derivative operator,

D2 (F ξ1 ) = D2 F ξ1 + 2DF Dξ1 + F D2 ξ1 ,

where
Dξ1 = h1 X (g1 ) + X (h1 ) g1 ,

D2 ξ1 = 2 (h1 g1 ) ,
and thus
D2 (F ξ1 ) = D2 F ξ1 + 2X (g1 ) (DF h1 ) + 2X (h1 ) (DF g1 )

+2F (h1 g1 ) = A1 + 2A2 + 2A3 + 2A4 .
Then,
E [hA1 , h2 g2 iH⊗2 ] = E[ξ1 D2 F, h2 g2 H⊗2 ]

= E[hD2 D2 F, h2 g2 H⊗2 , h1 g1 iH⊗2 ]

= E[ D4 F, h2 g2 h1 g1 H⊗4 ],

where we have applied Lemmas 4.3 and 4.1 in the second and third equalities
respectively. For the term B, we have
E [hA2 , h2 g2 iH⊗2 ]
= E [X (g1 ) hDF h1 , h2 g2 iH⊗2 ]
1 1
= [X (g1 ) hDF, h2 iH ] hh1 , g2 iH + [X (g1 ) hDF, g2 iH ] hh1 , h2 iH
2 2
1
2 1
= E[ D F, h2 g1 H⊗2 ] hh1 , g2 iH + E[ D2 F, g1 g2 H⊗2 ] hh1 , h2 iH .

2 2
Where we have used the integration by parts formula (2) and Lemma 4.1.
Analogously, for A3 we obtain
E [hA3 , h2 g2 iH⊗2 ]
1
1
= E[ D2 F, h1 h2 H⊗2 ] hg1 , g2 iH + E[ D2 F, h1 g2 H⊗2 ] hh2 , g1 iH .

2 2
Finally,
E [hA4 , h2 g2 iH⊗2 ] = E [F ] hh1 g1 , h2 g2 iH⊗2 .
Adding up all the terms the result follows.
Lemma 4.5. The exponential growth condition (4) implies (3).
Proof. The exponential growth assumption (4) implies

2 2
E[|∂ α f (Xt )| ] ≤ CT2 sup E[e2cT |Xt | ]. (6)
0≤t≤T
For any symmetric and positive definite matrix A we have

1/2
πd
Z
−hx,Axi
e dx = ,
Rd |A|
where |A| = det(A). As a consequence,
Z
2cT |Xt |2 1 1
E[e ]= d/2
e−hx,Axi dx = 1/2 1/2
,
(2π) |Vt | 1/2 Rd d/2
2 |Vt | |A|
with

1 1 −1
A = Vt−1 − 2cT Id = 2cT V − Id ,
2 4cT t
and this gives
2 −1/2
E[e2cT |Xt | ] = |Id − 4cT Vt | ,
provided A is symmetric and positive definite. This matrix is positive defi-

nite if and only if for all x ∈ Rd with |x| > 0

1 −1 1 T −1 2
xT Vt − Id x = x Vt x − |x| > 0,
4cT 4cT
which is implied by (5). Therefore,
h i
2 −1/2
E |∂ α f (Xt )| ≤ CT2 sup |Id − 4cT Vt | =: aT ,
0≤t≤T
which is finite by condition (5).
5. Convergence Results
From now on, C will denote a finite positive constant that may change from
line to line.
Proposition 5.1. Let

n−1
X d
X
R2π = 2
f (Xti ) ∆i X j ∆i X k − E ∆i X j ∆i X k .

∂j,k
i=0 j,k=1
Then
lim E[(R2π )2 ] = 0.
|π|→0
Proof. Set Fij,k = ∂j,k

2
f (Xti ) and
ϕj,k = ∆i X j ∆i X k − E ∆i X j ∆i X k = X(1jδi )X(1kδi ) − h1jδi , 1kδi iH .

i
Then
n−1 d
E[(R2π )2 ] = E[Fij11 ,k1 Fij22 ,k2 ϕji11 ,k1 ϕji22 ,k2 ],
X X
i1 ,i2 =0 j1 ,j2 ,k1 ,k2 =1
and by Lemma 4.4 we get the decomposition
E[Fij11 ,k1 Fij22 ,k2 ϕji11 ,k1 ϕji22 ,k2 ] = B1 + B2 + B3 + B4 + B5 + B6 ,
where
B1 = hE[D4 (Fij11 ,k1 Fij22 ,k2 )], 1jδ1i 1jδ2i 1kδi1 1kδi2 iH⊗4 ,
1 2 1 2
B2 = hE[D2 (Fij11 ,k1 Fij22 ,k2 )], 1jδ2i 1kδi1 iH⊗2 h1jδ1i , 1kδi2 iH ,
2 1 1 2
B3 = hE[D 2
(Fij11 ,k1 Fij22 ,k2 )], 1kδi1 1kδi2 iH⊗2 h1jδ1i , 1δj2i iH ,
1 2 1 2
B4 = hE[D2 (Fij11 ,k1 Fij22 ,k2 )], 1jδ1i 1jδ2i iH⊗2 h1kδi1 , 1kδi2 iH ,
1 2 1 2
B5 = hE[D2 (Fij11 ,k1 Fij22 ,k2 )], 1jδ1i 1kδi2 iH⊗2 h1jδ2i , 1kδi1 iH ,
1 2 2 1
B6 = 2E[Fij11 ,k1 Fij22 ,k2 ]h1jδ1i 1kδi1 , 1jδ2i 1kδi2 iH⊗2 .

1 1 2 2
Notice that the terms Bh , h = 1, . . . , 6, depend on the indices i1 , i2 , j1 , j2 ,

k1 , and k2 . We omit this dependence to simplify the notation and we set
n−1
X d
X
Bh = Bh ,
i1 ,i2 =0 j1 ,j2 ,k1 ,k2 =1
P6
so E[(R2π )2 ] = h=1 Bh . We have that
4
4
hE[Dp (Fij11 ,k1 ) D4−p (Fij22 ,k2 )], 1jδ1i 1jδ2i 1kδi1 1kδi2 iH⊗4 .
X
B1 =
p=0
p 1 2 1 2
On the other hand

p
p!
Dp (Fij,k ) =
X
2
∂ u (∂j,k f (Xti ))(11[0,ti ] )u1 · · ·(1d[0,ti ] )ud .
u1 ,...,ud =0
u1 ! · · · ud !
u1 +···+ud =p
Hence,
4 p 4−p
X 4 X X p! (4 − p)!
B1 =
p=0
p u1 ,...,ud =0 v1 ,...,vd =0
u1 ! · · · ud ! v1 ! · · · vd !
u1 +···+ud =p v1 +···+vd =4−p
E ∂ u (∂j21 ,k1 f (Xti1 ))∂ v (∂j22 ,k2 f (Xti2 )

×h(11[0,ti ] )u1 · · · (1d[0,ti ] )ud (11[0,ti ] )v1 · · · (1d[0,ti ] )vd ,

1 1 2 2
1jδ1i 1jδ2i 1kδi1 1kδi2 iH⊗4 .

1 2 1 2
Notice that
(11[0,ti ] )u1 · · · (1d[0,ti ] )ud (11[0,ti ] )v1 · · · (1d[0,ti ] )vd

1 1 2 2
= 1w w2 w3 w4
[0,s1 ] 1[0,s2 ] 1[0,s3 ] 1[0,s4 ] ,
1
where wk ∈ {1, . . . , d} , sk ∈ {ti1 , ti2 }, k = 1, . . . , 4. But for any 0 ≤ sk ≤

t, wk ∈ {1, . . . , d} , k = 1, . . . , 4,
j1 j2
|h1w w2 w3 w4 k1 k2
[0,s1 ] 1[0,s2 ] 1[0,s3 ] 1[0,s4 ] , 1δi1 1δi2 1δi1 1δi2 iH⊗4 |
1
1 X wσ(1) w w w
≤ |h1 0,s , 1j1 i h1 σ(2) , 1j2 i h1 σ(3) , 1k1 i h1 σ(4) , 1k2 i |
4! [ σ(1) ] δi1 H [0,sσ(2) ] δi2 H [0,sσ(3) ] δi1 H [0,sσ(4) ] δi2 H
σ∈Σ4
j1 j2 k1 k2
≤ sup |h1w w w w
[0,s] , 1δi iH ||h1[0,s] , 1δi iH ||h1[0,s] , 1δi iH ||h1[0,s] , 1δi iH |
1 2 1 2
0≤s≤t
1≤w≤d
= sup E[Xsw ∆i1 X j1 ] E[Xsw ∆i2 X j2 ] E[Xsw ∆i1 X k1 ] E[Xsw ∆i2 X k2 ] .

0≤s≤t
1≤w≤d
Furthermore, by Assumption (A4), we have
E[∂ u (∂j21 ,k1 f (Xti ))∂ v (∂j22 ,k2 f (Xti2 )] ≤ aT < ∞.

Hence, using Cauchy-Schwartz inequality,

 2
n−1
X d
X
sup E[Xsw ∆i X j ] E[Xsw ∆i X k ]

B1 ≤ CaT 

i=0 j,k=1 0≤s≤t
1≤w≤d
 2
d n−1
X X 2
sup E[Xsw ∆i X j ]  .

≤ CaT 

0≤s≤t
j,k=1 i=0
1≤w≤d
The last expression tends to zero as |π| → 0 by Assumption (A3). Analo-

gously
2 p 2−p
X 2 X X p! (2 − p)!
B2 =
p=0
p u ,...,u =0 v ,...,v =0 u1 ! · · · ud ! v1 ! · · · vd !
1 d 1 d
u1 +···+ud =p v1 +···+vd =2−p
E[∂ u (∂j21 ,k1 f (Xti1 ))∂ v (∂j22 ,k2 f (Xti2 )]

× h(11[0,ti ] )u1 · · · (1d[0,ti ] )ud (11[0,ti ] )v1 · · · (1d[0,ti ] )vd ,
1 1 2 2
1jδ2i 1kδi1 iH⊗2 h1jδ1i , 1kδi2 iH

2 1 1 2
≤ CaT sup E[Xsw ∆i2 X j2 ] E[Xsw ∆i1 X k1 ] E[∆i1 X j1 ∆i2 X k2 ] .

0≤s≤t
1≤w≤d
Therefore, by Cauchy-Schwartz inequality

 
n−1 d
X X 2 
sup E[Xsw ∆i X j ] 

B2 ≤ CaT 
i=0 j=1 0≤s≤t
1≤w≤d
 1/2
n−1 d
E[∆i1 X j ∆i2 X k ]2 
X X
×
i1 ,i2 =0 j,k=1
which tends to zero as |π| → 0 by Assumptions (A2) and (A3). The proof
for the terms B3 , B4 and B5 is almost the same as for the term B2 . Finally,
B6 = E[Fij11 ,k1 Fij22 ,k2 ]h1jδ1i , 1jδ2i iH h1kδi1 , 1kδi2 iH
1 2 1 2
+ E[Fij11 ,k1 Fij22 ,k2 ]h1jδ1i , 1kδi2 iH h1kδi1 , 1jδ2i iH

1 2 1 2
≤ aT E[∆i1 X j1 ∆i2 X j2 ] E[∆i1 X k1 ∆i2 X k2 ]

+ aT E[∆i1 X j1 ∆i2 X k2 ] E[∆i1 X k1 ∆i2 X j2 ] .

Hence,
 2
n−1
X d
X
B6 ≤ CaT  E[∆i1 X j ∆i2 X k ]
i1 ,i2 =0 j,k=1
d n−1
E[∆i1 X j ∆i2 X k ]2 ,
X X
≤ CaT
j,k=1 i1 ,i2 =0
which tends to zero as |π| → 0 by Assumption (A2).
Proposition 5.2. If
n−1
X d
X
R3π = 3
i=0 j,k,l=1
then
lim E[(R3π )2 ] = 0.
|π|→0
Proof. Setting
∆i X j ∆i X k ∆i X l = ∆i X j ∆i X k − E ∆i X j ∆i X k ∆i X l

+E ∆i X j ∆i X k ∆i X l ,

one gets
2
E[(R3π ) ]
 2 
n−1 d
 X X 3
∂j,k,l f (Xti ) ∆i X l ∆i X j ∆i X k − E ∆i X j ∆i X k  

≤ 2E 
i=0 j,k,l=1
 2 
n−1
X d
X
3
f (Xti ) ∆i X l E ∆i X j ∆i X k  

+2E  ∂j,k,l
 
i=0 j,k,l=1
= 2C1 + 2C2 .
To prove the convergence of C1 to zero, observe that
 2 
d
X n−1
X d
X n h io
3 l j k j k
C1 ≤ C E  ∂j,k,l f (Xti ) ∆i X ∆i X ∆i X − E ∆i X ∆i X  .
l=1 i=0 j,k=1
2
So it suffices to fix l and apply Proposition 5.1 with the term ∂j,k f (Xti )
3 l
replaced by ∂j,k,l f (Xti ) ∆i X =: g Xti , Xti+1 whose exact form does not
matter because it satisfies the exponential condition (4). Using Lemma 4.2,
we obtain that
n−1
X d
X
C2 = E[∂j31 ,k1 ,l1 f (Xti1 )∂j32 ,k2 ,l2 f (Xti2 )∆i1 X l1 ∆i2 X l2 ]
i1 ,i2 =0 j1 ,k1 ,l1 ,j2 ,k2 ,l2 =1
×E ∆i1 X j1 ∆i1 X k1 E ∆i2 X j2 ∆i2 X k2

= E1 + E2 ,
where Eh = in−1
P Pd
1 ,i2 =0 j1 ,k1 ,l1 ,j2 ,k2 ,l2 =1 Eh , for h = 1, 2, and
E1 = E[hD2 (∂j31 ,k1 ,l1 f (Xti1 )∂j32 ,k2 ,l2 f (Xti2 )), 1lδ1i 1lδi2 iH2 ]
1 2
j1 k1 j2 k2

×E ∆i1 X ∆i1 X E ∆i2 X ∆i2 X ,
E2 = E[∂j31 ,k1 ,l1 f (Xti1 )∂j32 ,k2 ,l2 f (Xti2 )]h1lδ1i , 1lδ2i iH
1 2
×E ∆i1 X j1 ∆i1 X k1 E ∆i2 X j2 ∆i2 X k2 .

Similarly to the preceding proposition, the term E1 can be bounded by

E1 ≤ CaT sup E[Xsw ∆i1 X l1 ] E[Xsw ∆i2 X l2 ]

0≤s≤t
1≤w≤d
×E ∆i1 X j1 ∆i1 X k1 E ∆i2 X j2 ∆i2 X k2 .

As a consequence, we obtain
  
n−1 d n−1 d
2
E ∆i X j ∆i X k 2  ,
X X X X
sup E[Xsw ∆i X j ]  

E1 ≤ CaT 

i=0 j=1 0≤s≤t i=0 j,k=1
1≤w≤d
where we have used the Cauchy-Schwartz inequality. This term tends to

zero as |π| → 0 by Assumptions (A2) and (A3). For the therm E2 , we have
E2 ≤ aT E ∆i1 X l1 ∆i2 X l2 E ∆i1 X j1 ∆i1 X k1 E ∆i2 X j2 ∆i2 X k2 ,

and by Cauchy-Schwartz inequality,

  
n−1 d n−1 d
E ∆i X k ∆j X l 2   E ∆i X j ∆i X k 2 
X X X X
E2 ≤ CaT 
i,j=0 k,l=1 i=0 j,k=1
which converges to 0 as |π| → 0 by Assumption (A2).
Proposition 5.3. Let X i be a point in the straight line that joins Xti and
Xti+1 and
n−1
X d
X
R4π = 4
∂j,k,l,m f (X i )∆i X j ∆i X k ∆i X l ∆i X m ,
i=0 j,k,l,m=1
then
lim E[ |R4π | ] =0.
|π|→0
Proof. We have
kR4π kL1 (Ω)
n−1
X d
X 4
∂j,k,l,m f (X i )∆i X j ∆i X k ∆i X l ∆i X m

≤ L1 (Ω)
i=0 j,k,l,m=1
n−1 d 1/2
2
X X
4
≤ E[(∂j,k,l,m f (X i )) ]
i=0 j,k,l,m=1
2 2 2 1/2
2
× E[ ∆i X j ∆i X k ∆i X l (∆i X m ) ] .
Appliying iteratively the Cauchy-Schwartz inequality one obtains

2 2 2 1/2
2
E[ ∆i X j ∆i X k ∆i X l (∆i X m ) ]
8 8 8 8
≤ (E[ ∆i X j ])1/8 (E[ ∆i X k ])1/8 (E[ ∆i X l ])1/8 (E[ (∆i X m ) ])1/8 .
Hence, by Assumption (A4)
 4
n−1 d
1/2
X X 8
kR4π kL1 (Ω) ≤ (aT )  (E[ ∆i X j ])1/8 

i=0 j=1
 4
n−1 d
1/2 j 2
X X
])1/2 

= (aT ) k (8) (E[ ∆i X
i=0 j=1
n−1 d
1/2 4
XX 2 2
≤ C (aT ) k (8) E[ ∆i X j ] (7)

i=0 j=1
where we have used that for all p > 0 if ξ is a centered Gaussian variable
one has
kξkLp (Ω) = κ (p) kξkL2 (Ω) ,
where
p+1
!1/p
√ Γ 2
κ (p) = 2 √ , p > 0.
π
And the last term in equation (7) converges to zero as |π| → 0 by Assump-
tion (A2).
6. Examples
6.1. Correlated Heterogeneous Fractional Brownian Motion
In this section we give an example where the theory previously developed
applies. Let B H = {(Bt1,H1 , . . . , Btd,Hd ), t ∈ [0, T ]} be a d-dimensional
heterogeneous fractional Brownian motion with Hurst parameter H =
d
(H1 , . . . , Hd ) ∈ (0, 1) and H1 ≤ · · · ≤ Hd . That is, B H is a d-dimensional
centered Gaussian process with covariance function matrix RH (s, t) given
by
i,j δij 2Hi 2H
RH (s, t) = δij RHi (s, t) = {s + t2Hi − |s − t| i },
2
for i, j = 1, . . . , d. Set Xt = ABtH , where A = (ai,j )i,j=1,...,d is a d×d matrix.
We call X a correlated heterogeneous fractional Brownian motion. X is
a d-dimensional Gaussian process, with the following correlation function

matrix
d d
" ! !#
E[Xsi Xtj ] aj,l Btl,Hl
X X
i,j
R (s, t) = =E ai,k Bsk,Hk
k=1 l=1
d
X
= ai,k aj,k RHk (s, t) .
k=1
Proposition 6.1. The process X = ABtH with 1/4 < min Hi < 1 satisfies
i
Assumptions (A1) to (A3). Therefore, the Wick-Itô formula applies to X.
Proof. We have that
d
Vtj,k = Rj,k (t, t) =
X
aj,m ak,m t2Hm ,
m=1
so Assumption (A1) is fulfilled. Let’s check Assumption (A2). For any k, l

we have
E ∆i X k ∆j X l

= Rk,l (ti+1 , tj+1 ) − Rk,l (ti+1 , tj ) − Rk,l (ti , tj+1 ) + Rk,l (ti , tj )
d
X
= ak,m al,m
m=1

k,l k,l k,l k,l
× RH m
(t , t
i+1 j+1 ) − R (t
Hm i+1 j, t ) − R (t
Hm i j+1, t ) + R (t
Hm i j , t )
d
1 X
= ak,m al,m
2 m=1

2H 2H 2H 2H
× |tj+1 − ti | m + |tj − ti+1 | m − |tj − ti | m − |tj+1 − ti+1 | m .
Therefore
n−1
E ∆i X k ∆j X l 2 = Bn + Cn + Dn ,
X
An =
i,j=0
where
n−1 d
!2
X X
Bn = ak,m al,m |ti+1 − ti |2Hm ,
i=0 m=1
n−2 d 2
!
1 X X
Cn = ak,m al,m |ti+2 − ti |2Hm − |ti+1 − ti |2Hm − |ti+2 − ti+1 |2Hm ,
2 i=0 m=1
n−3 n−1 d
1 X X X
Dn = ak,m al,m
2 i=0 j=i+2 m=1
!2

× |tj+1 − ti |2Hm + |tj − ti+1 |2Hm − |tj − ti |2Hm − |tj+1 − ti+1 |2Hm .
We have
n−1 d d
4Hm 4Hm −1
X X X
Bn ≤ C |ti+1 − ti | ≤ CT |π| ,
i=0 m=1 m=1
which converges to zero as |π| → 0 if H1 > 1/4. By a similar argument

we obtain the same result for Cn . The term Dn is more complicated. In
Nualart and Taqqu11 it is proved that, when j 6= i and j 6= i + 1,
2Hm 2Hm 2Hm 2Hm
|tj+1 − ti | + |tj − ti+1 | − |tj − ti | − |tj+1 − ti+1 |
2Hm −2 2Hm
≤ C (j − i − 1) |π| .
Then,
d n X
h
4Hm
X X
Dn ≤ C |π| k 4Hm −4 .
m=1 h=1 k=1
As n → +∞, one has the following asymptotics

4H −2

Xn X h  Cn m if Hm > 3/4
4Hm −4
k ∼ Cn ln n if Hm = 3/4 .
n if Hm < 3/4

h=1 k=1
−1
Since our partitions are in the class D we have n ≤ C |π| . Therefore,

2
 C |π|
 if H1 > 3/4
2 −1
Dn ≤ C |π| ln(|π| ) if H1 = 3/4 ,
 C |π|4H1 −1

if H1 < 3/4
and Assumption (A2) is fulfilled. Finally, let us check Assumption (A3).

One has
2 2
E Xtk ∆j X l = Rk,l (t, tj+1 ) − Rk,l (t, tj )

d
!2
1 X 2Hm 2Hm 2Hm 2Hm
= ak,m al,m (ti+1 − ti + |t − ti | − |t − ti+1 | )
4 m=1
d 2 2
2H 2H
X
≤C t2H 2Hm
i+1 − ti
m
+ |t − ti | m − |t − ti+1 | m .
m=1
2
As before, the convergence to zero of E Xtk ∆j X l

as |π| → 0 is con-
trolled by the term with H1 . If 1/4 < H1 ≤ 1/2, one has that t2H i+1 − ti
1 2H1
2H1 2H1 2H1
and |t − ti | − |t − ti+1 | are both bounded by (ti+1 − ti ) , and the
4H −1
sum of their squares is bounded by C |π| 1 , which converges to zero
as |π| → 0. If H1 > 1/2, either term is bounded by C (ti+1 − ti ) . Hence,
the sum of their squares is bounded by C |π| , which converges to zero as
|π| → 0. So the proof is concluded.
6.2. Multidimensional Fractional Brownian Motion

The d-dimensional fractional brownian motion with Hurst parameter H ∈
(0, 1) is the centered d-dimensional Gaussian process B H with the following
covariance function matrix RH (s, t)
δij 2H
i,j
RH (s, t) = δij RH (s, t) ={s + t2H − |s − t|2H }.
2
Obviously, B H is the process X considered in the previous section with
parameter (H, . . . , H) and A = Id , therefore we have the following result.
Proposition 6.2. The process B H with 1/4 < H < 1 satisfies Assump-
tions (A1) to (A3). Therefore, the Wick-Itô formula applies to B H .
7. Application to the Pricing of an Exchange Option

The market consists in two risky assets S 1 , S 2 and a risk free asset B.
Assume the following form for their dynamics
σ12 1,1

1 1 1
St = S0 exp µ1 t + σ1 Xt − V , S01 > 0,
2 t
σ22 2,2

2 2 2
St = S0 exp µ2 t + σ2 Xt − V , S01 > 0,
2 t
Bt = B0 exp {rt} , B0 > 0,
where X = X 1 , X 2 is the following correlated heterogeneous fractional

Brownian
Xt1 = Bt1,H1 ,
Xt2 = ρBt1,H1 + 1 − ρ2 Bt2,H2 ,
p
where ρ ∈ (0, 1) and H1 ≤ H2 . Note that

2
Vt1,1 = E[ Xt1 ] = t2H1 ,
2
Vt2,2 = E[ Xt2 ] = ρ2 t2H1 + 1 − ρ2 t2H2 ,

Vt1,2 = E[Xt1 Xt2 ] = ρt2H1 .

Suppose that H1 > 1/4, hence the Wick-Itô formula applies to X and we
obtain that
dSt1 = µ1 St1 dt + σ1 St1 dXt1 ,
dSt2 = µ2 St2 dt + σ2 St2 dXt2 .
+
Our aim is to price at time t ∈ [0, T ] the contingent claim ST1 − ST2 ,
which is known as an exchange option. Assume that the price process for
this option has the form C t, St1 , St2 , where C (t, x, y) is a function of class

C 1,2,2 and satisfies the exponential growth condition (4). Then the Wick-Itô
formula yields
Z t
∂C
C t, St1 , St2 = C 0, S01 , S02 + u, Su1 , Su2 du
0 ∂u
Z t Z t
∂C ∂C
+ µ1 u, Su1 , Su2 Su1 du + σ1 u, Su1 , Su2 Su1 dXu1
0 ∂x 0 ∂x
Z t Z t
∂C 1 2

2 ∂C
+ µ2 u, Su , Su Su du + σ2 u, Su1 , Su2 Su2 dXu2
0 ∂y 0 ∂y
Z t 2 2 0
1 2 ∂ C
+ σ1 2
u, Su1 , Su2 Su1 Vu1,1 du
2 0 ∂x
1 2 t ∂2C 2
2,2 0
Z
1 2 2
+ σ2 2
u, S u , S u S u V u du
2 0 ∂y
Z t 2 0
∂ C
+ σ1 σ2 u, Su1 , Su2 Su1 Su2 Vu1,2 du. (8)
0 ∂x∂y
The price C t, St1 , St2 should coincide with the value at time t of a port-

+
folio which replicates the contingent claim ST1 − ST2 . Let Πt denote the
amuount of this portfolio invested in the risk free asset Bt an h1t , h2t the
amount of stocks S 1 and S 2 , respectively. Then,
C t, St1 , St2 = Πt + h1t St1 + h2t St2 .

We will consider portfolios satisfying the following Wick self-financing type

condition
Z t
C t, St1 , St2 = C 0, S01 , S02 + r

Πu du
0
Z t Z t
+ µ1 h1u Su1 du + σ1 (h1u Su1 ) dXu1 (9)
0 0
Z t Z t
+ µ2 h2u Su2 du + σ2 (h2u Su2 ) dXu2 .
0 0
RT
We also suppose that the portfolio is admissible, that is, 0 |Πt | dt <
R T i i i
∞, 0 ht dt < ∞ and hu Su is Wick forward integrable on any inter-
val [0, t], i = 1, 2. Choosing ht = ∂x t, St , St and h2t = ∂C
1 ∂C 1 2 1 2

∂y t, St , St and
comparing equations (8) and (9) we get that C (t, x, y) must satisfy the
partial differential equation
∂C ∂C ∂C
rC = +r x+r y
∂t ∂x ∂y
σ 2 ∂ 2 C 2 1,1 0 σ22 ∂ 2 C 2 2,2 0 ∂2C
+ 1 x (Vt ) + y (Vt ) + σ1 σ2 xy(Vt1,2 )0 , (10)
2 ∂x2 2 ∂y 2 ∂x∂y
with terminal condition
+
C (T, x, y) = (x − y)
and boundary conditions
C (t, 0, y) = 0,
C (t, x, 0) = x.
8
Reasoning as Margrabe, C (t, x, y) is homogeneous of degree 1 in x and y.
Therefore, thanks to Euler’s theorem for homogeneous functions, we have
that
∂C (t, x, y) ∂C (t, x, y)
C (t, x, y) = x +y
∂x ∂y
and equation (10) simplifies to
∂C σ 2 ∂ 2 C 2 1,1 0 σ22 ∂ 2 C 2 2,2 0 ∂2C
+ 1 x (Vt ) + y (Vt ) + σ1 σ2 xy(Vt1,2 )0 = 0.
∂t 2 ∂x2 2 ∂y 2 ∂x∂y
Using again the homogeneity of C (t, x, y) we can define C(t, z) :=
C (t, x, y) /y where z = x/y and find the following partial differential equa-
tion for C
∂C z2 ∂2C
+ {σ12 (Vt1,1 )0 + σ22 (Vt2,2 )0 − 2σ1 σ2 (Vt1,2 )0 } 2 = 0, (11)
∂t 2 ∂z
with terminal condition

+
C (T, z) = (z − 1)
and boundary condition
C (t, 0) = 0.
Define
θ (t) := σ12 (Vt1,1 )0 + σ22 (Vt2,2 )0 − 2σ1 σ2 (Vt1,2 )0 ,
then the solution to equation (11) is
C (t, z) = zN (d1 ) − N (d2 ),
where
RT s
ln z + 21 θ (s) ds T
Z
t
d1 := qR , d2 := d1 − θ (s) ds,
T t
t θ (s) ds
and N (x) is the N (0, 1) cumulative distribution function.
Finally, taking into account the values of Vt1,1 , Vt2,2 and Vt1,2 , we get
C t, St1 , St2 = St1 N (d1 ) − St2 N (d2 ) ,

where d1 and d2 are obtained from d1 and d2 making z = St1 /St2 and
Z T
θ (s) ds = σ12 + σ22 − 2ρσ1 σ2 T 2H1 − t2H1

t
+ σ22 1 − ρ2 T 2H2 − t2H2 .

References
1. E. Alòs, O. Mazet and D. Nualart, Stochastic calculus with respect to Gaus-
sian processes, Ann. Probab. 29, 766–801 (2001).
2. P. Carmona and L.Coutin, Stochastic integration with respect to fractional
Brownian motion, Ann. Inst. H. Poincaré 39, 27–68 (2003).
3. L. Decreusefond and A. S. Üstünel, Stochastic analysis of the fractional Brow-
nian motion, Potential Analysis 10, 177–214 (1998).
4. T. E. Duncan, Y. Hu and B. Pasik-Duncan, Stochastic calculus for fractional
Brownian motion I. Theory, SIAM J. Control Optim. 38, 582–612 (2000).
5. Y. Hu, Integral transformations and anticipative calculus for fractional Brow-
nian motions, Memoirs of the AMS 175 (2005).
6. S. Janson, Gaussian Hilbert Spaces (Cambridge University Press, Cambridge,
1997).
7. B. B. Mandelbrot and J. W.Van Ness, Fractional Brownian motions, frac-
tional noises and applications, SIAM Review 10, 422–437 (1968).
8. W. Margrabe, The value of an option to exchange one asset for another, The
Journal of Finance 33, 177-186 (1978).
9. D. Nualart, Stochastic integration with respect to fractional Brownian motion
and applications, Contemporary Mathematics 336, 3–39 (2003).
10. D. Nualart, The Malliavin Calculus and Related Topics, (Springer-Verlag,
Berlin, 2006).
11. D. Nualart and M.S. Taqqu, Wick-Itô formula for regular processes and appli-
cations to the Black and Scholes formula, Stochastics and Stochastics Reports,
to appear.
12. D. Nualart and M.S. Taqqu, Wick-Itô formula for Gaussian processes, J.
Stoch. Anal. Appl. 24, 599–614 (2006).
13. T, Sottinen, Fractional Brownian motion in finance and queueing, Ph.D.
Thesis, University of Helsinki (2003).
May 11, 2011 10:58 WSPC - Proceedings Trim Size: 9in x 6in 02-tsoi
27
Fractional White Noise Multiplication
Allanus H. Tsoi
Department of Mathematics, University of Missouri
Columbia, MO 65211, USA
Email: tsoia@missouri.edu
Fractional Brownian motion can be viewed as a generalization of the classical

Brownian motion process, which can be described through the associated Hurst
parameter. The Hurst parameter is a description of the memory of the frac-
tional Brownian motion. The range of the Hurst parameter is between 0 and 1.
The classical Brownian motion is a special version of the fractional Brownian
motion when the Hurst parameter equal 1/2. In this paper we study fractional
Brownian motion in terms of the standard white noise, with Hurst parameter
1
2
< H < 1. We consider fractional white noise as the fractional integral of the
standard white noise. We construct a fractional differential operator and study
some of its properties. Finally we formulate the notion of fractional white noise
multiplication.
1. Introduction
White noise multiplication has been discussed in the book by H.-H. Kuo.12
One of its advantage over the Ito differential is that it generalizes the dif-
fusion coefficient term b(t, Xt ) in the classical Ito differential equation:
dXt = a(t, Xt )dt + b(t, Xt )dBt , (1)
in the sense that the b(t, Xt )dBt is replaced by the white noise term Ḃt ·
b(t, Xt )dt in such a way that the term b comes from a comparatively much
wider class of functions.
The application of white noise calculus to solve a class of Cauchy prob-
lems were presented in the work by Chung, Ji and Saitô.3 On the other
hand, the classical investment and consumption problem can be formu-
lated and solved in terms of solving a Cauchy problem, as can be seen in
the book by I. Karatzas and S.E. Shreve.11 As a consequence, our fractional
white noise multiplication results can be used to: (1) generate and solve a
wide class of fractional Cauchy problems, which, in turn, (2) can be applied
28 A. H. Tsoi
to solve the fractional white noise counterpart of the classical investment

and consumption problem.
A fractional Brownian motion is an extension of the notion of the clas-
sical Brownian motion, in the sense that a fractional Brownian motion
is a Gaussian process which possesses “memory”. The study of fractional
Brownian motion started in the sixties, when Mandelbrot and Van Ness
published their paper in SIAM.14 Ever since then applied mathematicians
and econometricians have been employing the concept of fractional Brow-
nian motion to study and analyse financial and economic data in the case
when their data exhibit autocorrelation phenomena.
In this paper we study fractional Brownian motion in terms of white
noise analysis. Physically speaking, white noise is the time derivative of
Brownian motion. Since the sample paths of a Brownian motion is of un-
bounded variation, we need to study the derivative of the sample paths of
Brownian motion in terms of generalized function theory. The analysis of
Brownian motion sample paths in terms of generalized function theory is
called white noise analysis. The first study of white noise analysis appeared
in the work of T. Hida.6 Ever since then numerous applications of white
noise analysis on quantum physics and mathematical finance emerged. See
Refs. 2, 14, and 18. In this paper we first present some background of the
standard Gaussian white noise calculus which can be found in Hida, Kuo,
Potthoff and Streit,8 Kuo,12 or Saitô and Tsoi.19
Consider the Gel’fand triple:
S(R) ⊂ L2 (R) ⊂ S 0 (R) (2)
on the real line R. We use (L2 ) = L2 (S 0 (R), µB ) to denote the standard

white noise space of square integrable functionals. We recall the Wiener-Itô
theorem which says that any function φ ∈ (L2 ) can be decomposed uniquely
as a sum of multiple Wiener integrals:
∞
X
φ= In (f n ), f n ∈ L̂2c (Rn ), (3)
n=0
where In is the multiple Wiener integral of order n with respect to the

Brownian motion, and L̂2c (Rn ) denotes the space of symmetric complex-
valued L2 -funtions on Rn .
The (L2 )-norm ||φ||0 of φ is given by:
∞
!1/2
X
n 2
||φ||0 = n!|f |0 , (4)
n=0
Fractional White Noise Multiplication 29
where | · |0 denotes the L2c (Rn )-norm for any n.

Let A denote the operator A = −(d/du)2 + u2 + 1. For φ given by the
expression Eq. (3) above, and which satisfies the condition:
∞
X
n!|A⊗n f n |20 < ∞, (5)
n=0
we define Γ(A)φ ∈ (L2 ) by:

∞
X
Γ(A)φ = In (A⊗n f n ). (6)
n=0
The operator Γ(A) is called the second quantization operator of A.

For each p ≥ 0, define
||φ||p = ||Γ(A)p φ||0 , (7)
where || · ||0 is the (L2 )-norm. Let
(Sp ) = {φ ∈ (L2 ) : ||φ||p < ∞}. (8)
Then (Sp ) is a Hilbert space with norm || · ||p . Define
(S) = projective limit of {(Sp ); p ≥ 0}. (9)
Then (S) is a nuclear space, which is called the space of test functions. The
dual space (S)0 of (S) is called a space of generalized functions (or Hida
distributions). Thus we have the Gel’fand triple:
(S) ⊂ (L2 ) ⊂ (S)0 . (10)
0
The bilinear pairing of (S) and (S) will be denoted by hh·, ·ii.
The S-transform SΦ of a generalized function Φ ∈ (S)0 is defined to be
the function:
SΦ(ξ) = hhΦ, : eh·,ξi :ii, ξ ∈ Sc (R), (11)
or equivalently,
1
SΦ(ξ) = e− 2 hξ,ξi hhΦ, eh·,ξi ii. ξ ∈ Sc (R). (12)
Here Sc (R) denotes the complexification of S(R).
If Φ(x) = ∞ ⊗n
:, F n i ∈ (S)0 , then its S-transform is given by:
P
n=0 h: x
∞
X
SΦ(ξ) = hF n , ξ ⊗n i, ξ ∈ Sc (R). (13)
n=0
Further properties of the S-transform can be found in Hida et.al.8 or

Kuo.12
30 A. H. Tsoi
We note that similar results hold if we replace R by the half-line

(−∞, T ], where T > 0 is a fixed positive constant throughout this paper.
The fractional Brownian motion (FBM) B H (t), t ∈ R, with Hurst pa-
rameter H, 0 < H < 1, and starting value B H (0) = 0, is given by:
Z 0
H 1
B (t) = { [(t − v)H−1/2 − (−v)H−1/2 ]dB(v)
Γ(H + 12 ) −∞
Z t (14)
+ (t − v)H−1/2 dB(v)},
0
for t > 0, and is defined similarly for t < 0.

See Ref. 14 for the above definitions of fractional Brownian motions.
FBM has received much attention in a variety of applications for several
decades. See, for example, Refs. 1, 2, 4, 13, 14, 15, and 18. When 12 < H < 1,
FBM has a long range dependence in the sense that if we let
r(n) = cov(B H (1), (B H (n + 1) − B H (n))), (15)
then
∞
X
r(n) = ∞. (16)
n=1
In this paper we study the fractional Brownian motion with Hurst param-
eter 21 < H < 1 in terms of the standard Gaussian white noise. We express
the FBM as a fractional integral of Gaussian white noise in the sense of
generalized functionals. We construct a fractional differential operator in
Sec. 3 and study some of its properties. In Sec. 4 we formulate the no-
tion of fractional white noise multiplication, while in Sec. 5 we present an
application of the above concept on a Binomial type finance model.
2. Fractional Integral Representation of Fractional

Brownian Motion
We recall some results about generalized funtions according to Gel’fand and
Shilov.5
First we discuss the direct product of two generalized functions f (x) and
g(y) of one variable defined on the space of infinitely differentiable functions
with bounded supports. Consider a test function φ(x, y), (x, y) ∈ R2 . The
direct product f (x) × g(y) is the generalized function on the space of test
functions φ(x, y) defined as
< f (x) × g(y), φ(x, y) >=< f (x), < g(y), φ(x, y) >> . (17)
Now suppose that both the generalized functions f (x) and g(y) have
support on the same interval (−∞, T ] for some constant T > 0. Consider
(f (x), φ(x + y)) where φ has bounded support, which is an infinitely differ-
entiable function of y. For sufficiently large negative values of y, the support
of φ(x + y) does not intersect with that of f (x). Thus for such y the func-
tion (f (x), φ(x + y)) vanishes, and its support is therefore bounded on the
left. But the support of g(y) is bounded on the right by assumption. Hence
for large enough a > 0, the strip |x + y| ≤ a that contains the support of
φ(x + y) has a bounded intersection with the support of the product f × g.
Thus in |x + y| ≤ a we can replace φ(x + y) by a function φ(x, y) with
bounded support having the same values in the intersection of the strip
and the support of f × g. Now we can define the convolution f ∗ g of two
such generalized functions f and g by:
< f ∗g, φ >=< f (x)×g(y), φ(x+y) >=< g(y), < f (x), φ(x+y) >> . (18)
Similarly, if the supports of both f and g are bounded on the left side (i.e.,
f = 0 for x < a and g = 0 for y < b), then the product f × g and the
convolution f ∗ g can be defined. See page 103, Gel’fand and Shilov.5
Now we consider generalized functions g concentrated on the interval
(−∞, T ], for fixed constant T > 0. Define, for λ > 0,
γTλ = xλ
if x ∈ (−∞, T ], and
γTλ (x) = 0
otherwise.
Next we define the fractional integral of g of order λ > 0 as the convo-
lution:
γTλ−1
Iλ (g)(x) = gλ (x) = (g ∗ )(x). (19)
Γ(λ)
Note that in the above definition, the function γTλ−1 is locally summable.
γ λ−1
If λ ≤ 0, we have to view Γ(λ) T
as a generalized function, and regulariza-
tion and normalization are involved. See Chapters 3 and 5 in Gel’fand and
Shilov.5
Since Eq. (19) can also be expressed as:
Z t
1
Iλ (g)(t) = (t − v)λ−1 g(v)dv, (20)
Γ(λ) −∞
32 A. H. Tsoi
we see that if g is an ordinary continuous function with bounded support,

and if λ = n is a positive integer, then Eq. (20) is just the usual n-folded
integral of g given by the Cauchy formula (see Ref. 5, pg. 115) or through
integration-by-parts.
Next, for λ > 0 the fractional derivative of order λ of g is defined as the
convolution
γT−λ−1
Dλ (g)(x) = g ∗ . (21)
Γ(−λ)
We also use the convention that, if λ > 0, we write
I−λ (g) = Dλ (g),
and
D−λ (g) = Iλ (g).
Proposition 2.1. For α > 0 and λ > 0 and generalized function g with
support (−∞, T ], we have:
Iα Iλ (g)) = Iα+λ (g). (22)
Proof.
Z t Z z
1 1
Iα (Iλ (g))(t) = (t − z)α−1 (z − v)λ−1 g(v)dvdz
Γ(α) −∞ Γ(λ) −∞
Z t Z t
1
= g(v) (t − z)α−1 (z − v)λ−1 dzdv.
Γ(α)Γ(λ) −∞ v
Make a change of variable w = z − v so that the above quantity equals

Z t Z t−v
1
g(v) (t − w − v)α−1 wλ−1 dwdv.
Γ(α)Γ(λ) −∞ 0
w
Next let u = t−v so that the above integral becomes
Z t Z 1
1
(t − v)α+λ−1 g(v) (1 − u)α−1 uλ−1 dudv
Γ(α)Γ(λ) −∞ 0
which is the same as

Z t
1
(t − v)α+λ−1 g(v)dv = Iα+λ (g)(t),
Γ(α + λ) −∞
since
Z 1
Γ(α)Γ(λ)
(1 − u)α−1 uλ−1 du = .
0 Γ(α + λ)
See also Ref. 16, page 72.
For the rest of this paper we shall concentrate on the Hurst parameter
H, with 12 < H < 1. Recall that the fractional Brownian motion(FBM)
with Hurst parameter H, 12 < H < 1, can be expressed in the form given
on the right hand side of Eq. (14).
Now we consider the white noise space (S 0 (R), µB ), where µB is the
standard white noise measure and B is the standard Brownian motion
given in Sec. 1.
Denote Ḃ(t)(x) = dB(t)(x)
dt = x(t) as the sample path of the Brownian
noise (time derivative of the Brownian motion) evaluated at the sample
path x ∈ S 0 (R).
We would like to express B H as
Z t
B H (t) = Ḃ H (τ )dτ (23)
0
for some generalized function Ḃ H ∈ S 0 (R).

According to R. Barton,2 let
Z t
1 3
Ḃ H (t) = |t − τ |H− 2 Ḃ(τ )dτ
Γ(H − 12 ) −∞ (24)
= IH− 12 (Ḃ)(t).
If we substitute this into Eq. (23), then, upon changing the order of
integration, Eq. (23) becomes
1
Z t Z τ 3
B H (t) = |τ − s|H− 2 Ḃ(s)dsdτ
Γ(H − 12 ) 0 −∞
1
Z 0 Z t 3
Z t Z t 3

= |τ − s|H− 2 dτ Ḃ(s)ds + |τ − s|H− 2 dτ Ḃ(s)ds
Γ(H − 12 ) −∞ 0 0 s
1
Z 0 1 1
Z t 1

= |t − s|H− 2 − |s|H− 2 Ḃ(s)ds + |t − s|H− 2 Ḃ(s)ds
Γ(H + 12 ) −∞ 0
which agrees with the definition of fractiona Brownian motion given in

expression Eq. (14). We call Ḃ H in Eq. (24) the fractional white noise of
Hurst parameter H.
34 A. H. Tsoi
We also have
Z tZ τ
1 3
B H (t) = 1 |τ − s|H− 2 Ḃ(s)dsdτ
Γ(H − 2 ) 0 −∞
Z t
= Ḃ H (τ )dτ
0
Z t
= IH− 12 (Ḃ)(τ )dτ
0
= hχ[0,t] , IH− 12 (Ḃ)i
Z
= χ[0,t] (τ )IH− 12 (Ḃ)(τ )dτ.
R
Remark 2.1.
xH (t) = hxH , δt i
= Ḃ H (t)(x) (25)
= IH− 12 (x)(t).
Rt
Note that in the above Eq. (25) the integral 0
IH− 12 (s) is understood to be
hIH− 12 (Ḃ), I[0,t] i, which is the generalized function IH− 21 (Ḃ) ∈ S ∗ acting
the L( R) function I[0,t] , which is interpreted as the L2 (S ∗ , µ) limit of a
sequence hIH− 21 (Ḃ), ξn i, where the sequence {ξn } ⊂ S(R) converging in
L2 (R) to I[0,t] .
3. Fractional Differential Operator
Definition 3.1. For y ∈ S 0 ((−∞, T ]), Φn (x) = h: x⊗n :, f n i ∈ (S), with

f n ∈ L̂2c,CS (Rn ) denoting functions in L̂2c (Rn ) with compact supports,
define the fractional differential operator as:
DH
y Φn (x) := nh: x
⊗(n−1)
:, hyH , f n ii, (26)
where
Z
n
hyH , f i(·) = f (·, t)yH (t)dt ∈ L2C (Rn−1 ), (27)
R
and
yH = IH− 12 (y) ∈ S 0 (R). (28)

Consequently,
∞
X 1
DH
y (: e
h·,ξi
:) = n h: x⊗(n−1) :, hyH , ξiξ ⊗(n−1) i
n=1
n! (29)
h·,ξi
= hξ, yH i : e :,
and
∂tH (: eh·,ξi :) = DδHt (: eh·,ξi :)

= hξ, δtH i : eh·,ξi : .
By definition,
∗ ∗
S(DH H
y Φ)(ξ) = hhDy Φ, : e
h·,ξi
:ii
= hhΦ, DyH (: eh·,ξi :)ii (30)
= hξ, yH iSΦ(ξ).
That is,
∗
DH
y Φ(x) = hx, yH iΦ(x). (31)
Thus
∗
DH
y (h: x
⊗n
:, f n i) = hx, yH ih: x⊗n :, f n i
(32)
= h: x⊗(n+1) :, f n ⊗y
ˆ H i.
Remark 3.1.
∗
∂tH Φ(x) = hx, δtH iΦ(x). (33)
In particular,
∗
∂tH (h: x⊗n :, f n i) = hx, δtH ih: x⊗n :, f n i
(34)
= h: x⊗(n+1) :, f n ⊗δ
ˆ tH i.
Proposition 3.1.
hx, yH i = hxH , yi. (35)
In particular,
hx, δtH0 i = hxH , δt0 i.

36 A. H. Tsoi
Proof. Note that

Z
hx, yH i = x(t)IH− 12 y(t)dt
R
Z Z t
1 3
= x(t) 1 (t − v)H− 2 y(v)dvdt
R Γ(H − ) −∞
Z Z ∞ 2
3
= y(v) (t − v)H− 2 x(t)dtdv.
R v
However, letting w = −(t − v), we have

Z ∞ Z −∞
H− 32 3
(t − v) x(t)dt = (−w)H− 2 x(v − w)d(−w)
v 0
Z 0
3
=− wH− 2 x(v − w)dw
∞
Z v
3
= (v − u)H− 2 x(u)du,
−∞
where the last equality is obtained by letting u = v − w. Thus from the

above, we have
Z Z v
3
hx, yH i = y(v) (v − u)H− 2 x(u)dudv
R −∞
= hIH− 21 (x), yi
= hxH , yi.
Remark 3.2. For φ =: eh·,ξi :, ψ =: eh,·,ηi :,

DνH φ(x) = hνH , ξi(: eh·,ξi :); (36)
DνH ψ(x) = hνH , ηi(: eh·,ηi :). (37)

Thus
hhDνH φ, ψii + hhφ, DνH ψii = hhhνH , ξi : eh·,ξi :, ψii + hhφ, hνH , ηiψii
= hνH , ξ + ηihh: eh·,ξi :, : eh·,ηi :ii (38)
= hνH , ξ + ηiehη,ξi .
4. Fractional White Noise Multiplication
Definition 4.1. For Φn = h: x⊗n :, f n i, f n ∈ S(Rn ) with compact sup-

port, define
KH Φn := h: x⊗n :, fH
n
i, (39)
where
n (n)
fH (t1 , . . . , tn ) = IH− 1 (t1 , . . . , tn )
2
Z t1 Z tn
3 3
= ... (t1 − v1 )H− 2 . . . (tn − vn )H− 2 f n (v1 , . . . , vn )dvn . . . dv1 .
−∞ −∞
(40)
Definition 4.2. For Φ ∈ (S)0 , φ, ψ ∈ (S), define the fractional multiplica-
tion φ ◦H Φ of order H by:
hhφ ◦H Φ, ψii := hhφ(KH Φ), ψii
(41)
= hhKH Φ, φψii.
Remark 4.1. From Definition 4.1, for Φ = h·, νi and ν ∈ S(R) with
compact support, we obtain KH Φ(x) = h·, νH i.
Next, for φ =: eh·,ξi :, ψ =: eh·,ηi :, where ξ, η ∈ S(R), we have
hhh·, νi ◦H φ, ψii = hhh·, νH i, φψii
Z
1 2 2
= e− 2 (|ξ|0 +|η|0 ) h·, νH ieh·,ξ+ηi dµ(x) (42)
= hνH , ξ + ηiehη,ξi ,
where the last equality is a consequence of Lemma 9.16 on page 113 from
Kuo.12
Theorem 4.1. The fractional white noise multiplication is given by:
∗
Ḃ H (t) ◦H φ = (∂tH + ∂tH )φ, φ ∈ (S). (43)
Proof. First by Eq. (28),

hhh·, νi ◦H φ, ψii = hhDνH φ, ψii + hhφ, DνH ψii. (44)
Now,
hhh·, νi ◦H φ, h: ·⊗n :, f n iii = hhh·, νH ih: ·⊗n :, f n i, φii. (45)
12
Hence by Lemma 9.19 on page 114 of Kuo, together with Eq. (26) and
Eq. (32), we have
h·, νH ih: ·⊗n :, f n i = h: ·⊗n : ⊗x, ˆ ni
ˆ νH ⊗f
= nh: x⊗(n−1) : ⊗τ, ˆ n i + h: x⊗(n+1) :, νH ⊗f
ˆ νH ⊗f ˆ ni
= nh: x⊗(n−1) :, hνH , f n ii + h: x⊗(n+1) :, νH ⊗f
ˆ ni
∗
= DνH h: ·⊗n :, f n i + DνH h: ·⊗n :, f n i.
(46)
38 A. H. Tsoi
Hence
hhh·, νi ◦H φ, h: ·⊗n : f n iii = hhh·, νH ih: ·⊗n :, f n i, φii
∗
= hhDνH h: ·⊗n :, f n i + DνH h: ·⊗n :, f n i, φii
∗
= hh(DνH + DνH )φ, h: ·⊗n :, f n iii
(47)
which implies
∗
h·, νi ◦H φ = (DνH + DνH )φ. (48)
In particular, for
Ḃ H (t)(x) = hxH , δt i,
we have
Ḃ H (t)(x) = hx, δtH i
∗
= DδHt + DδHt (49)
∗
= ∂tH + ∂tH .
That is,
∗
Ḃ H (t) ◦H φ = (∂tH + ∂tH )φ. (50)
5. A Binomial-Type Finance Model

Consider an economy with only two states: the upstate and the downstate,
respectively. The probability operator of the upstate is:
2
ḂH,u
p= 2 2
(51)
ḂH,u + ḂH,d
and the probability operator of the downstate is:
2
ḂH,d
p= 2 2
, (52)
ḂH,u + ḂH,d
where ḂH,u and ḂH,d are interpreted as fractional white noise operater
with parameter H as discussed in the previous sections, and we define the
2 2
squares ḂH,u and ḂH,d as:
2 ∗
ḂH,u ξ = ((∂H,u,· + ∂H,u,· )ξ)2 ; (53)
and
2 ∗
ḂH,d ξ = ((∂H,d,· + ∂H,d,· )ξ)2 . (54)
There are two securities: one risk-free bond with interest rate R and one
stock with initial price S(0) and time t = 1 price S(1) with u and d being
the returns of the stock at the upstate and at the downstate at time t = 1.
A trading strategy for a portfolio is a pair (B0 , ∆0 ) where B0 is the dollar

amounts in the bond and the ∆0 the number of shares of the stock at time
t = 0. Thus the value of the portfolio at times t = 0 and t = 1 are:
V (0) = B0 + ∆0 S(0) (55)

and
V (1) = B0 (1 + R) + ∆0 S(1). (56)
The price system of the bond and the stock admits arbitrage if and only if
there is a trading strategy (B0 ), ∆0 ) with V (0) = B0 + ∆0 S(0) = 0 such
that V (1) ≥ 0 and P (V (1) > 0) > 0. It can be shown that the price system
does not admit arbitrage if and only if d < 1 + R < u. Suppose that the
condition d < 1 + R < u is violated, say, 1 + R ≤ d. It is easy to verify
that ∆0 = 1 and B0 = −S(0) is an arbitrage trading strategy. The case
u ≤ 1 + R is argued similarly. Conversely, if d < 1 + R < u, then for any
strategy (B0 , ∆0 ) with V (0) = B0 + ∆0 S(0),
V (1) = B0 (1 + R) + ∆0 S(1), (57)

and is equal to ∆0 S(0)[−(1 + R) + u] if in the upstate, and is equal to
∆0 S(0)[−(1 + R) + d] if in the downstate, implying that V (1) cannot be
non-negative. In other words, there is no-arbitrage strategy. In the following
we assume that the price system does not admit arbitrage and derive option
prices using a no-arbitrage argument.
40 A. H. Tsoi
Consider a European option which gives its holder the right (not obliga-
tion) to receive or pay a pre-specified amount that is contingent on the
state of the economy on a specific date. For example, a Euorpean call op-
tion written on a risky security gives its holder the right (not obligation)
to buy the underlying security at a pre-specified price on a specific date;
while a European put option written on a security gives its holder the right
(not obligation) to sell the underlying security at a pre-specified prie on
a specific date. The pre-specified price is called the strike price and the
specific date is called the expiration or maturity time. The payoff functions
for a European call option and a European put option are maxS(T ) − K, 0
and maxK − S(T ), 0, respectively, where K is the strike price and T is the
expiration time.
Consider now that there is an option which pays Cu at the upstate and Cd
at the downstate at time t = 1. To determine the price of this option, we
construct a portfolio hoping that the payoff of this portfolio is the same as
that of the option. If we can do so, then the price of the option must be
equal to the initial value of the portfolio in order to avoid arbitrage. For
the option with payoff Cu or Cd , the corresponding portfolio is constructed
as follows. The trading strategy (B0 , ∆0 ) should be such that the following
payoff equations are satisfied:
B0 (1 + R) + ∆0 S(0)u = Cu (58)
and
B0 (1 + R) + ∆0 S(0)d = Cd . (59)
Solving these equations yields
Cu − Cd
∆0 S(0) = , (60)
u−d
and
uCd − dCu
B0 = . (61)
(1 + R)(u − d)
Therefore the price of the option is given by
B0 + ∆0 S(0) = (1 + R)−1 [pCu + (1 − p)Cd ], (62)

where
2
ḂH,u 1+R−d
p= 2 2
= . (63)
ḂH,u + ḂH,d u−d
Therefore we have
2 1+R−d 2
ḂH,u = Ḃ . (64)
u − 1 − R H,d
Note that the right hand side of the above equation is intepreted as an
operator defined on a suitable domain of generalized functions.
References
1. E. Alos, O. Mazet and D. Nualart: Stochastic calculus with respect to Gaus-
sian Processes, Ann. of Probability, Vol. 29, No. 2 (2001), p. 766-801.
2. R.J. Barton: Signal Detection in Fractional Gaussian Noise, IEEE Transac-
tions on Information Theory, Vol. 34, No. 5, (1988), p. 943-959.
3. D.M. Chung, U.C. Ji and K. Saitô: Cauchy problems associated with the Lévy
Laplacian in white noise analysis. Infinite dimensional Analysis, Quantum
Probability and Related Topics, Vol. 2, No.1 (1999). 131-153.
4. T.E. Duncan, Y. Hu and B. Pasik-Duncan: Stochastic calculus for fractional
Brownian motion. I. Theory. SIAM Journal of Control and Optimization 38,
no.2 (2000), 582- 612.
5. I.M. Gel’fand and G.E. Shilov: Generalized Functions. Vol. I. Academic Press
1964.
6. T. Hida: Canonical representations of Gaussian processes and their applica-
tions. Memoirs Coll. Sci., Univ. Kyoto A33 (1960), 109-155.
7. T. Hida and M. Hitsuda: Gaussian Processes. American Math. Soc. Trans-
lation of Math. Monographs Vol. 12, 1993.
8. T. Hida, H.-H. Kuo, J. Potthoff and L. Streit: White Noise Analysis. Kluwer,
1993.
9. H. Holden, B. Oksendal, J. Uboe and T. Zhang: Stochastic Partial Differential
Equations. Birkhauser 1996.
10. Y. Hu: Ito-Wiener chaos expansion with exact residual and correlation, vari-
ance inequalities. Journal of Theoretical Probability, 10 (1997), 835- 848.
11. I. Karatzas and S.E. Shreve: Brownian Motion and Stochastic Calculus. Sec-
ond Edition. Springer-Verlag, New York 1991.
12. H.-H. Kuo: White Noise Distribution Theory. CRC Press 1996.
13. S.J. Lin: Stochastic analysis of fractional Brownian motion. Stochastics and
Stochastic Reports 55 (1995), 121- 140.
14. B.B. Mandelbrot and J.W. Van Ness: Fractional Brownian motion, fractional
noises and applications. SIAM Rev. 10 (1968), 422- 437.
42 A. H. Tsoi
15. K. Nishi, K. Saitô and A.H. Tsoi: Fractional Brownian motion and the Lévy
Laplacian. Quantum Information V. World Scientific (2006), p. 181-191.
16. I. Podlubny: Fractional Differential Equations. Academic Press, 1999.
17. P. Protter: Stochastic Integration and Differential Equations. Springer- Ver-
lag, 1990.
18. L.C.G. Rogers: Arbitrage with fractional Brownian motion. Math. Finance
7 (1997), 95-105.
19. K. Saitô and A.H. Tsoi: The Levy Laplacian as a self- adjoint operator.
Quantum Information. Ed. T. Hida and K. Saitô, World Scientific (1999),
159- 171.
May 11, 2011 11:3 WSPC - Proceedings Trim Size: 9in x 6in 03-chao
43
Invariance Principle of Regime-Switching Diffusions
C. Zhu
Department of Mathematical Sciences
University of Wisconsin-Milwaukee
Milwaukee, WI 53201, USA
Email: zhu@uwm.edu
G. Yin
Department of Mathematics
Wayne State University
Detroit, MI 48202, USA
Email: gyin@math.wayne.edu
This work is devoted to a two-component process termed switching diffusions.

The underlying process has a continuous component with diffusive behavior
and a discrete component responsible for the movement from one discrete
regime to another. Different from the Markov-modulated switching diffusions,
the discrete component depends on the continuous component. Invariance prin-
ciples of LaSalle are derived under suitable conditions.
Keywords: Switching diffusion; stability; invariance theorem.
1. Introduction
1.1. Introduction
This work is concerned with large-time behavior of switching diffusion pro-
cesses, which are two-component Markov processes. In such processes, be-
cause of the two components, continuous dynamics and discrete events co-
exist and are intertwined. The models are versatile and much enlarge the
domain of applicability of diffusion-type processes. Our motivation of the
study stems from a wide variety of applications in financial engineering,
production planning, wireless communications, and other related fields, in
which the traditional stochastic differential equation formulation is not ad-
equate. In addition to the usual dynamic systems modeled by differential
equations, there is another factor in the model that could be used to cap-
ture the environmental changes and other random factors. In lieu of one
44 C. Zhu and G. Yin
dynamic system represented by a differential equation, we encounter a finite

number of such equations for different regimes. Across the different regimes,
the behavior of the system could be drastically different. For example, one
of the early efforts of using such hybrid models for financial applications
can be traced back to Ref. 1, in which both the appreciation rate and the
volatility of a stock depend on a continuous-time Markov chain. The intro-
duction of regime-switching models makes it possible to describe stochastic
volatility in a relatively simple manner. To illustrate, in the simplest case, a
stock market may be considered to have two “modes” or “regimes,” up and
down, resulted from the state of the underlying economy, the general mood
of investors in the market, and so on. The rationale is that in the different
modes or regimes, the volatility and return rates are very different.
The rapid progress in natural science, life science, engineering, as well
as in social science demands the consideration of such systems. In fact, the
advent of switching diffusions is largely because of the practical needs in
modeling complex dynamic systems. In recent year, there have been much
research activities in studying such systems; see Refs. 2,5,9,12,15,21,23,25,
28,29. Some most recent development on switching diffusion processes can
be found in Ref. 24 and references therein. For systems running for a long
time, it is crucial to have definitive characterization of such systems’ long
run behavior; see Refs. 12,15,25 for recent progress on stability of such
systems.
1.2. Our Contributions

Most of the work to date has been concentrated on Markov-modulated dif-
fusions, in which the Brownian motion and the switching force are indepen-
dent, whereas less is known for systems with continuous-component (termed
x-dependent henceforth for simplicity of presentation) dependent switching
processes. As demonstrated in Ref. 29, when x-dependent switching dif-
fusions are encountered, even such properties as continuous and smooth
dependence on the initial data are nontrivial and are fairly difficult to es-
tablish. Nevertheless, studying such systems are both practically useful and
theoretically interesting. In our recent work, basic properties such as recur-
rence, positive recurrence, and ergodicity are studied in Ref. 28; stability
is treated in Ref. 12; stability of randomly switching ordinary differential
equations is treated in Ref. 30. Continuing in this direction, this paper takes
up the issue of examination of invariance principle akin to the LaSalle’s the-
orem for deterministic systems Refs. 6,7,14. It should be mentioned that
Invariance Principle of Regime-Switching Diffusions 45
study of invariance principles for stochastic systems can also be found in

Refs. 13,16 etc.
1.3. Applications to Mathematical Finance

This work will be of interest to researchers working in the area of mathe-
matical finance. Early work on regime-switching models for option pricing
can be found in Ref. 1. Subsequent work in economic time series Ref. 8 ex-
tends much of the traditional ARMA model setup. As this time, there have
been resurgent interests in using regime-switching diffusions to depict the
financial market, where the switching or jump processes are used to describe
stochastic volatility resulting from market modes, interest rates, as well as
other economic factors. Some recent work in this direction can be found in
Refs. 2,3,18,26,27 and many references therein. In addition to finite-time
horizon properties, it is desirable to learn large-time behavior of various
models including the well-know geometric Brownian motion models, the
regime-switching models, and mean-reversion models. In this regard, vari-
ous notions of stability including stability in probability as well as almost
surely, and stability in distribution are of considerable interests. In related
insurance and risk models, one is interested in the ruin probability and
possibility of default etc., where regime-switching models have also gained
popularity (see, for example, Ref. 20). These consideration include finite
time analysis and large-time behavior, which is related to stability as well.
In this work, we aim to establish invariance principles for regime-switching
diffusions, which will help us further in understanding the underpinning
of scenarios arising in financial problems. It will open up new domains for
further investigation on issues including stability, mean reversion, ergodic
measures, and impact on mathematical finance and related issues.
1.4. Outline
The rest of the paper is arranged as follows. Section 2 begins with the for-
mulation of the problem. Section 3 is devoted to the invariance principles.
Liapunov function type criteria are obtained first. Then linear (in x) sys-
tems are treated. Finally, a few more remarks are made in Section 4 to
conclude the paper.
2. Formulation
Throughout the paper, we use z 0 to denote the transpose of z ∈ R`1 ×`2
with ì ≥ 1, whereas R`×1 is simply written as R` ; 11 = (1, 1, . . . , 1)0 ∈ Rm0
is a column vector with all entries being 1; the Euclidean norm for a row
or a column vector x is denoted by |x|. As usual, I denotespthe identity
matrix. For a matrix A, its trace norm is denoted by |A| = tr(A0 A). If
a matrix A is real and symmetric, we use λmax (A) and λmin (A) to denote
the maximal and minimal eigenvalues of A, respectively, and set ρ(A) :=
max {|λmax (A)| , |λmin (A)|}. When B is a set, IB (·) denotes the indicator
function of B.
Let (Ω, F , P) be a complete probability space. Suppose that (X(t), α(t))
is a two-component Markov process such that X(·) is a continuous compo-
nent taking values in Rr and α(·) is a jump component taking values in a
finite set M = {1, 2, . . . , m0 }. The process (X(t), α(t)) has a generator L
given as follows. For each i ∈ M and any twice continuously differentiable
function g(·, i),
r r
1 X ∂ 2 g(x, i) X ∂g(x, i)
Lg(x, i) = ajk (x, i) + bj (x, i) + Q(x)g(x, ·)(i),
2 ∂xj ∂xk j=1
∂xj
j,k=1
1
= tr(a(x, i)∇2 g(x, i)) + b0 (x, i)∇g(x, i) + Q(x)g(x, ·)(i),
2
(1)
where x ∈ Rr , and Q(x) = (qij (x)) is an m0 × m0 matrix depending on x
P
satisfying qij (x) ≥ 0 for i 6= j and j∈M qij (x) = 0,
X X
Q(x)g(x, ·)(i) = qij (x)g(x, j) = qij (x)(g(x, j)−g(x, i)), i ∈ M,
j∈M j∈M,j6=i
and ∇g(·, i) and ∇2 g(·, i) denote the gradient and Hessian of g(·, i), respec-
tively.
The process (X(t), α(t)) can be described by
dX(t) = b(X(t), α(t))dt + σ(X(t), α(t))dw(t), X(0) = x, α(0) = α, (2)
and for i 6= j
P{α(t + ∆t) = j|α(t) = i, (X(s), α(s)), s ≤ t} = qij (X(t))∆t + o(∆t), (3)
where w(t) is a d-dimensional standard Brownian motion, b(·, ·) : Rr ×M 7→

Rr , and σ(·, ·) : Rr × M 7→ Rr×d satisfies σ(x, i)σ 0 (x, i) = a(x, i). Note
that the evolution of the discrete component α(·) can be represented as
a stochastic integral with respect to a Poisson random measure (see, for
example, Refs. 5 and 19). Indeed, for x ∈ Rr and i, j ∈ M with j 6= i,
let ∆ij (x) be consecutive (with respect to the lexicographic ordering on
M × M), left closed, right open intervals of the real line, each having
length qij (x). Define a function h : Rr × M × R 7→ R by

m0
X
h(x, i, z) = (j − i)I{z∈∆ij (x)} . (4)
j=1
Then (3) is equivalent to

Z
dα(t) = h(X(t), α(t−), z)p(dt, dz), (5)
R
where p(dt, dz) is a Poisson random measure with intensity dt × m(dz),
and m is the Lebesgue measure on R. The Poisson random measure p(·, ·)
is independent of the Brownian motion w(·). Denote the natural filtration
by Ft := σ {(X(s), α(s)), s ≤ t}. Without loss of generality, assume the
filtration {Ft }t≥0 satisfies the usual condition (i.e., it is right continuous
with F0 containing all P-null sets).
Throughout the paper, we assume that both b(·, i) and σ(·, i) satisfy
the usual Lipschitz condition and linear growth condition for each i ∈ M
and that Q(·) is bounded and continuous. It is well known that under these
conditions, the system (2)–(3) (or equivalently, (2)–(5)) has a unique strong
solution; see Refs. 10 or 22 for details. From time to time, we often wish to
emphasize the initial data (X(0), α(0) = (x, α) dependence of the solution
of (2)–(3), which will be denoted by (X x,α (t), αx,α (t)) similar to its diffusion
counter part.
A few words on notations are needed at this point. We define for any
(x, i), (y, j) ∈ Rr × M that
(
|x − y|, if i = j
d((x, i), (y, j)) =
|x − y| + 1, if i 6= j.
It is easy to verify that for any (x, i), (y, j), and (z, l) we have (i)
d((x, i), (y, j)) ≥ 0 and d((x, i), (y, j)) = 0 if and only if (x, i) = (y, j), (ii)
d((x, i), (y, j)) = d((y, j), (x, i)), and (iii) d((x, i), (y, j)) ≤ d((x, i), (z, l)) +
d((z, l), (y, j)). Thus d is a distance function of Rr ×M. Also if U is a subset
of Rr × M, we define
d((x, i), U ) = inf {d((x, i), (y, j)) : (y, j) ∈ U } .
Let M be endowed with the trivial topology. Then it is obvious that for a
fixed U ∈ Rr × M, the function d(·, U ) is continuous. As usual, we denote
by d(x, D) the distance between x ∈ Rr and D ⊂ Rr , that is
d(x, D) = inf {|x − y| : y ∈ D} .
Throughout the paper, we use Px,α and Ex,α to denote the probability and
expectation with (X(0), α(0)) = (x, α), respectively.
3. Main Results
This section is the core of the paper. We begin with the definition of in-
variant sets. Inspired by the study in Ref. 4, we define the invariant set as
follows.
Definition 3.1. A Borel measurable set U ⊂ Rr ×M is said to be invariant

with respect to the solutions of (2)–(5) or simply, U is invariant with respect
to the process (X(t), α(t)) if
Px,i {(X(t), α(t)) ∈ U, for all t ≥ 0} = 1, for any (x, i) ∈ U,
that is, a process starting from in U will remain in U a.s.
As shown in Refs. 12 and 15, when the coefficients of (2)–(5) satisfy

b(0, α) = σ(0, α) = 0, for each α ∈ M,
then almost every sample path of any solution with initial condition (x, i)
satisfying x 6= 0 will never reach the origin, in other words, the set (Rr −
{0}) × M is invariant with respect to the solutions of (2)–(5).
Using the terminologies in Refs. 4 and 11, we present the definitions of
stability and asymptotic stability. Then general results in terms of Liapunov
function are provided.
Definition 3.2. A closed and bounded set K ⊂ Rr × M is said to be
(i) stable in probability if for any ε > 0 and ρ > 0, there is a δ > 0
such that

Px,i sup d((X(t), α(t)), K) < ρ ≥ 1 − ε, if d((x, i), K) < δ;
t≥0
(ii) asymptotically stable in probability is it is stable in probability, and

moreover
n o
Px,i lim d((X(t), α(t)), K) = 0 → 1, as d((x, i), K) → 0;
t→∞
(iii) stochastically asymptotically stable in the large if it is stable in prob-

ability, and
n o
Px,i lim d((X(t), α(t)), K) = 0 = 1, for any (x, i) ∈ Rr × M;
t→∞
(iv) asymptotically stable with probability one if

lim d((X(t), α(t)), K) = 0, a.s.
t→∞
Theorem 3.3. Assume that there exists a nonnegative function V : Rr ×

M 7→ R+ such that Ker(V ) := {(x, i) ∈ Rr × M : V (x, i) = 0} is nonempty
and bounded, and that for each α ∈ M, V (·, α) is twice continuously dif-
ferentiable with respect to x, and
LV (x, i) ≤ 0, for all (x, i) ∈ Rr × M. (1)
Then
(i) Ker(V ) is an invariant set for the process (X(t), α(t)), and
(ii) Ker(V ) is stable in probability.
Proof. Let (x0 , i0 ) ∈ Ker(V ). By virtue of the generalized Itô Lemma (Ref.
19), we have for any t ≥ 0,
Z t
V (X(t), α(t)) = V (x0 , i0 ) + LV (X(s), α(s))ds + M (t), (2)
0
where M (t) = M1 (t) + M2 (t) is a local martingale with

Z t

M1 (t) = ∇V (X(s), α(s)), σ(X(s), α(s))dw(s) ,
0
Z tZ

M2 (t) = V (X(s), i0 + h(X(s), α(s−), z))
0 R

− V (X(s), α(s)) µ(ds, dz),

where ·, · denotes the usual inner product, µ(ds, dz) = p(ds, dz) − ds ×
m(dz) is a martingale measure (p(dt, dz) is the Poisson random measure
with intensity dt × m(dz) as in (5)). Taking expectations on both sides of
(2) (as in Ref. 28, use a sequence of stopping times and Fatou’s lemma, if
necessary), it follows from (1) that
Ex0 ,i0 [V (X(t), α(t))] ≤ V (x0 , i0 ) = 0.
But V is nonnegative, so we must have V (X(t), α(t)) = 0 a.s. for any t ≥ 0.
Then we have
( )
Px0 ,i0 sup V (X(tn ), α(tn )) = 0 = 1,
tn ∈Q+
where Q+ denotes the set of nonnegative rational numbers. Now by virtue

of Ref. 29, Lemma 3.6, the process (X(t), α(t)) is cádlág (sample paths
being right continuous and having left limits). Thus we obtain

Px0 ,i0 sup V (X(t), α(t)) = 0 = 1.
t≥0
That is
Px0 ,i0 {(X(t), α(t)) ∈ Ker(V ), for all t ≥ 0} = 1.
This proves the first assertion of the theorem.
We proceed to prove the second assertion. For any δ > 0, let Uδ be a
neighborhood of Ker(V ) such that
Uδ := {(x, i) ∈ Rr × M : d((x, i), Ker(V )) < δ} . (3)
Let the initial condition (x, i) ∈ Uδ − Ker(V ) and τ be the first exit time
of the process from Uδ . That is,
τ = inf{t : (X(t), α(t)) 6∈ Uδ }.
Then for any t ≥ 0, we have by virtue of generalized Itô’s lemma that
Z t∧τ
V (X(t ∧ τ ), α(t ∧ τ )) = V (x, i) + LV (X(s), α(s))ds + M (t ∧ τ ),
0
where
Z t∧τ

M (t ∧ τ ) = ∇V (X(s), α(s)), σ(X(s), α(s))dw(s)
0Z
t∧τ Z
+ V (X(s), i + h(X(s), α(s−), z))
0 R
−V (X(s), α(s)) µ(ds, dz).
As argued in the previous paragraph, by virtue of (1), we can use a sequence
of stopping times and Fatou’s Lemma, if necessary, to obtain
Z t∧τ
Ex,i [V (X(t∧τ ), α(t∧τ ))] ≤ V (x, i)+Ex,i LV (X(s), α(s))ds ≤ V (x, i).
0
Since V is nonnegative, we further have

V (x, i) ≥ Ex,i [V (X(τ ), α(τ ))I{τ <t} ] + Ex,i [V (X(t), α(t))I{t≤τ } ]
(4)
≥ Ex,i [V (X(τ ), α(τ ))I{τ <t} ].
For notational simplicity, denote (ξ, `) = (X(τ ), α(τ )). We claim that
V (ξ, `) > ρ, for some constant ρ > 0. (5)
Sk
To this end, write Ker(V ) = l=1 (Njl × {jl }), where k ≤ m0 , Njl ⊂ Rr
and jl ∈ M, for jl = 1, . . . , k. Further, we denote J = {j1 , . . . , jk } ⊂ M.
For example, with Rr = R, m0 = 3, V (x, 1) = x2 , V (x, 2) = (x + 1)2 , and
V (x, 3) = 1 + x2 . Then Ker(V ) = ({0} × {1}) ∪ ({−1} × {2}) or N1 = {0}
and N2 = {−1}.
Let us first consider the case when ` 6∈ J. Note that ξ ∈ D, where D is a

Sk
bounded neighborhood of l=1 Nl (such a neighborhood D exists because
Ker(V ) is bounded by assumption). Then we have
inf {V (x, `) : x ∈ D} ≥ ρ1 > 0. (6)
Suppose (6) were not true. Then there would exist a sequence {xn } ⊂ D
such that lim V (xn , `) = 0. Since {xn } is bounded, there exists a sub-
n→∞
sequence {xnk } such that xnk → x
e. Thus by the continuity of V (·, `), we
have
V (e
x, `) = lim V (xnk , `) = 0.
k→∞
x, `) ∈ Ker(V ). This is a contradiction to the assumption that

That is, (e
` 6∈ J. Thus (6) is true and hence V (ξ, `) ≥ ρ1 .
Now let us consider the case when ` ∈ J. It follows that δ ≤ d(ξ, N` ) ≤
δe < ∞. A similar argument using contradiction as in the previous case
shows that
n o
inf V (x, `) : δ ≤ d(x, N` ) ≤ δe ≥ ρ2 > 0.
Thus it follows that V (ξ, `) ≥ ρ2 . A combination of the two cases gives us

V (ξ, `) ≥ ρ, where ρ = ρ1 ∧ ρ2 , Hence the claim follows.
Finally, we have from (4) and (5) that
1
Px,i {τ < t} ≤ V (x, i).
ρ
Letting t → ∞,
1
Px,i {τ < ∞} ≤ V (x, i).
ρ
Note that

{τ < ∞} = sup d((X(t), α(t)), Ker(V )) ≥ δ .
0≤t<∞
Therefore it follows that

1
Px,i sup d((X(t), α(t)), Ker(V )) ≥ δ ≤ V (x, i) → 0,
0≤t<∞ ρ
as d((x, i), Ker(V )) → 0. This finishes the proof of the theorem. 2
Next we consider asymptotic stability. To this end, we need the following
lemma.
Lemma 3.4. Assume that there exists a nonnegative function V : Rr ×

M 7→ R+ with nonempty and bounded Ker(V ), such that for each α ∈ M,
V (·, α) is twice continuously differentiable with respect to x, and that for

any ε > 0
LV (x, i) ≤ −κε < 0, for any (x, i) ∈ (Rr × M) − U ε , (7)
where κε is a positive constant depending on ε, Uε is a neighborhood of
Ker(V ) as defined in (3), and U ε denotes the closure of Uε . Then for any
0 < ε < r, we have
Px,i {τε,r < ∞} = 1, for any (x, i) ∈ Uε,r ,
where
Uε,r = {(y, j) ∈ Rr × M : ε < d((y, j), Ker(V )) < r} ,
and τε,r is the first exit time from Uε,r , that is
τε,r := inf {t ≥ 0 : (X(t), α(t)) 6∈ Uε,r } .
Proof. Fix any (x, i) ∈ Uε,r . By virtue of generalized Itô’s Lemma, we have
that for any t ≥ 0,
Z t∧τε,r
V (X(t∧τε,r ), α(t∧τε,r )) = V (x, i)+ LV (X(s), α(s))ds+M (t∧τε,r ),
0
where M (t ∧ τε,r ) is a martingale with mean zero. Thus by taking expecta-

tions on both sides, and using (7), we obtain
Z t∧τε,r
Ex,i [V (X(t ∧ τε,r ), α(t ∧ τε,r ))] ≤ V (x, i) − Ex,i κε ds
0
= V (x, i) − κε Ex,i [t ∧ τε,r ].
Note that V is nonnegative, hence we have
1
Ex,i [t ∧ τε,r ] ≤ V (x, i).
κε
But
Ex,i [t ∧ τε,r ] = tPx,i {τε,r > t} + Ex,i [τε,r I{τε,r ≤t} ] ≥ tPx,i {τε,r > t} .
Thus it follows that
1
tPx,i {τε,r > t} ≤ V (x, i).
κε
Now letting t → ∞, we have
Px,i {τε,r = ∞} = 0 or Px,i {τε,r < ∞} = 1.
The assertion thus follows. 2
Theorem 3.5. Assume that there exists a function V satisfying the con-
ditions of Lemma 3.4. Then Ker(V ) is an invariant set for the process
(X(t), α(t)) and Ker(V ) is asymptotically stable in probability.
Proof. Our proof is motivated by Ref. 17, Theorem 5.36; we use similar
ideas. By virtue of Theorem 3.3, we know that Ker(V ) is an invariant set
for the process (X(t), α(t)) and that Ker(V ) is stable in probability. Hence
it remains to show that
n o
Px,i lim d((X(t), α(t)), Ker(V )) = 0 → 1 as d((x, i), Ker(V )) → 0.
t→∞
Since Ker(V ) is stable in probability, for any ε > 0 and any θ > 0, there
exists some δ > 0 (without loss of generality, we may assume that δ < θ)
such that

ε
Px,i sup d((X(t), α(t)), Ker(V )) < θ ≥ 1 − , (8)
t≥0 2
for any (x, i) ∈ Uδ , where Uδ is defined in (3). Now fix any (x, α) ∈ Uδ −
Ker(V ) and r > d((x, i), Ker(V )). let ρ > 0 be arbitrary satisfying 0 < ρ <
d((x, i), Ker(V )) and choose some % ∈ (0, ρ). Define
τ% := inf {t ≥ 0 : d((X(t), α(t)), Ker(V )) ≤ %} ,

τθ := inf {t ≥ 0 : d((X(t), α(t)), Ker(V )) ≥ θ} .
Then it follows from Lemma 3.4 that
Px,α {τ% ∧ τθ < ∞} = Px,α {τ%,θ < ∞} = 1, (9)
where τ%,θ is the first exit time from U%,θ , and
U%,θ := {(y, j) ∈ Rr × M : % < d((X(t), α(t)), Ker(V )) < θ} .
But (8) implies that Px,α {τθ < ∞} ≤ 2ε . Note also
Px,α {τ% ∧ τθ < ∞} ≤ Px,α {τ% < ∞} + Px,α {τθ < ∞} .
Thus it follows that

ε
Px,α {τ% < ∞} ≥ Px,α {τ% ∧ τθ < ∞} − Px,α {τθ < ∞} ≥ 1 − . (10)
2
Now let
τρ := inf {t ≥ τ% : d((X(t), α(t)), Ker(V )) ≥ ρ} .

(We use the convention that inf {∅} = ∞.) For any t ≥ 0, we apply gener-
alized Itô’s lemma and (7) to obtain
Ex,α V (X(τρ ∧ t), α(τρ ∧ t))
Z τρ ∧t
≤ Ex,α V (X(τ% ∧ t), α(τ% ∧ t)) + Ex,α LV (X(s), α(s))ds (11)
τ% ∧t
≤ Ex,α V (X(τ% ∧ t), α(τ% ∧ t)).

Note that τ% ≥ t implies τρ ≥ t and hence on the set {ω ∈ Ω : τ% (ω) ≥ t}
we have
Ex,α V (X(τρ ∧ t), α(τρ ∧ t)) = Ex,α V (X(t), α(t))
(12)
= Ex,α V (X(τ% ∧ t), α(τ% ∧ t)).
Hence it follows from (11) and (12) that

Ex,α I{τ% <t}
V (X(τρ ∧ t), α(τρ ∧ t))
≤ Ex,α I{τ% <t} V (X(τ% ∧ t), α(τ% ∧ t))

= Ex,α I{τ% <t} V (X(τ% ), α(τ% )) ≤ Vb% ,
where Vb% := sup {V (y, j) : d((y, j), Ker(V )) = %}. Note that τρ < t implies
τ% < t. Hence we further have

Vb% I{τ% <t}V (X(τρ ∧ t), α(τρ ∧ t))
≥ Ex,α I{τ% <t} I{τρ <t} V (X(τρ ∧ t), α(τρ ∧ t))
= Ex,α I{τρ <t} V (X(τρ ∧ t), α(τρ ∧ t))
= Ex,α I{τρ <t} V (X(τρ ), α(τρ ))
≥ Vρ Px,α {τρ < t} ,
where Vρ := inf {V (y, j) : ρ ≤ d((y, j), Ker(V )) ≤ ρe}, with ρe > 0 being
some constant. Recall that we showed in the proof of Theorem 3.3 that
Vρ > 0. Now since V is continuous, we choose % sufficiently small so that
Vb% ε
Px,α {τρ < t} ≤ ≤ .
Vρ 2
Letting t → ∞, we obtain
ε
Px,α {τρ < ∞} ≤ . (13)
2
Finally, it follows from (10) and (13) that
Px,α {τ% < ∞, τρ = ∞} ≥ Px,α {τ% < ∞} − Px,α {τρ < ∞} ≥ 1 − ε.
This implies that

Px,α lim sup d((X(t), α(t)), Ker(V )) ≤ ρ ≥ 1 − ε.
t→∞
But ρ > 0 can be chosen to be arbitrarily small. Therefore we have

n o
Px,α lim d((X(t), α(t)), Ker(V )) = 0 ≥ 1 − ε.
t→∞
This finishes the proof of the theorem. 2
Theorem 3.6. Assume there exists a function V satisfying the conditions

of Lemma 3.4. If V also satisfies
lim inf V (x, α) = ∞. (14)
|x|→∞ α∈M
Then Ker(V ) is asymptotically stable in probability in the large, that is,

Ker(V ) is stable in probability and
n o
Px,α lim d((X(t), α(t)), Ker(V )) = 0 = 1, (15)
t→∞
r
for any (x, α) ∈ R × M.
Proof. By virtue of Theorem 3.3, Ker(V ) is stable in probability. Thus it

remains to verify (15). To this end, as in the proof of Theorem 3.3, write
Ker(V ) = kl=1 (Njl × {jl }), where k ≤ m0 , Njl ⊂ Rr , and jl ∈ M. Since
S
Sk
Ker(V ) is bounded by assumption, in particular, l=1 Njl is bounded, there
exists some R > 0 such that
[k
sup |y| : y ∈ Njl ≤ R. (16)
l=1
r
Let ε > 0 and fix any (x, α) ∈ R × M. Then (14) implies that there exists
some positive constant β > (R + 2) ∨ d((x, α), Ker(V )) such that
2V (x, α)
inf {V (y, j) : |y| ≥ β, j ∈ M} ≥ . (17)
ε
Define
τβ := inf {t ≥ 0 : d((X(t), α(t)), Ker(V )) ≥ 2β} .
For any t ≥ 0, we have by virtue of generalized Itô’s lemma and (7) that
Ex,α V (X(t ∧ τβ ), α(t ∧ τβ )) ≤ V (x, α). (18)
We claim that |X(τβ )| ≥ β. If this were not true, it would follow from (16)
that for any (y, j) ∈ Ker(V ),
d((X(τβ ), α(τβ )), (y, j)) ≤ |X(τβ ) − y| + 1 < β + R + 1 < 2β − 1,
where in the last inequality above, we used the fact that β > R + 2. Then
we have d((X(τβ ), α(τβ )), Ker(V )) ≤ 2β − 1 < 2β. This is a contradiction
with the definition of τβ . Thus we must have |X(τβ )| ≥ β. Then it follows

from (18) that

V (x, α) ≥ Ex,α V (X(τβ ), α(τβ ))I{τβ <t}
≥ inf {V (y, j) : |y| ≥ β, j ∈ M} Px,α {τβ < t} ,
and hence (17) implies that

ε
Px,α {τβ < t} ≤ .
2
By letting t → ∞, we have
ε
Px,α {τβ < ∞} ≤ .
2
Then we can finish the proof using the same argument in the proof of
Theorem 3.5. 2
The following theorem provides a criterion for asymptotic stability with
probability 1.
Theorem 3.7. If there exists a nonnegative function V : U 7→ R+ such that

for each α ∈ M, V (·, α) is twice continuously differentiable with respect to
c : Rr × M 7→ R+ satisfying
x and that there exists a continuous function W
LV (x, i) ≤ −W
c (x, i), for any (x, i) ∈ U, (19)
where U ⊂ Rr ×M is an invariance set for the process (X(t), α(t)). Assume

also that either U is bounded or
lim V (x, α) = ∞. (20)

|x|→∞, (x,α)∈U
Then for any initial condition (x, α) ∈ Rr × M, the following assertions

are true
(i) lim sup V (X x,α (t), αx,α (t)) < ∞ a.s.,

t→∞
(ii) Ker(W
c ) 6= ∅,
(iii) lim d((X x,α (t), αx,α (t)), Ker(W
c )) = 0 a.s., and
t→∞
(iv) if moreover, Ker(W
c ) = {0} × M, then
lim X x,α (t) = 0 a.s. .

t→∞
Proof. This theorem can be proved using the arguments in Ref. 16, Theo-
rem 2.1. Some modifications are needed, though. 2
We end this section with the following results on linear systems. More
specifically, we assume that the evolution (2) is replaced by
d
X
dX(t) = b(α(t))X(t)dt + σj (α(t))X(t)dwj (t), (21)
j=1
where b(i), σj (i) are r × r constant matrices and wj (t) are independent
1-dimensional standard Brownian motions for i = 1, 2, . . . , m0 and j =
1, 2, . . . , d.
Note that 0 is an equilibrium point for the system given by (21) and
(3). As we indicated earlier, it was shown in Refs. 12 and 15 that the set
Rr × M is invariant with respect to the process (X(t), α(t)).
Theorem 3.8. Assume that the discrete component α(·) is ergodic with
constant generator Q = (qij ) and invariant distribution π = (π1 , . . . , πm0 ) ∈
R1×m0 . Then the equilibrium point x = 0 of system given by (21) and (3)
(i) is asymptotically stable with probability 1 if
m0
X X d
0 0
πi λmax b(i) + b (i) + σj (i)σj (i) < 0, (22)
i=1 j=1
(ii) is unstable in probability if

Xm0 d
X
πi λmin b(i) + b0 (i) + σj (i)σj0 (i)
i=1 j=1 (23)

1 2
− ρ(σj (i) + σj0 (i)) > 0.
2
Proof. We need only prove assertion (i), since assertion (ii) was proved
in Ref. 12. For notational simplicity, define the column vector µ =
(µ1 , µ2 , . . . , µm0 )0 ∈ Rm0 with
d
1 0
X
0
µi = λmax b(i) + b (i) + σj (i)σj (i) .
2 j=1
Also let β := −πµ. Note that β > 0 by (22). It follows from the result in
Ref. 12 that the equation
Qc = µ + β11
has a solution c = (c1 , c2 , . . . , cm0 )0 ∈ Rm0 . Thus we have
m0
X
µi − qij cj = −β, i ∈ M. (24)
j=1
For each i ∈ M, consider the Liapunov function V (x, i) = (1 − γci )|x|γ ,

where 0 < γ < 1 is sufficiently small so that 1 − γci > 0 for each i ∈ M.
It is readily seen that for each i ∈ M, V (·, i) is continuous, nonnegative,
vanishes only at x = 0, and satisfies (20). Detailed calculations as in the
proof of Ref. 12, Theorem 4.3, reveal that for x 6= 0, we have

 x0 b(i)x X m0
γ cj − ci
LV (x, i) = γ(1 − γci )|x| − qij
 |x|2 1 − γci
j=1
! (25)
d 0 0 0 0 2 
1 X x σj (i)σj (i)x (x σj (i)x)
+ + (γ − 2) .
2 j=1
|x|2 |x|4 
Note that
d
x0 b(i)x 1 X x0 σj0 (i)σj (i)x
+
|x|2 2 j=1 |x|2
d
x0 (b0 (i) + b(i))x 1 X x0 σj0 (i)σj (i)x
= + (26)
2|x|2 2 j=1 |x|2
d
1 0
X
0
≤ λmax b(i) + b(i) + σj (i)σj (i) = µi .
2 j=1
Next, it follows that when γ is sufficiently small,

m0 m0 m0
X cj − ci X X ci (cj − ci ) X
qij = qij cj + qij γ= qbij cj + O(γ). (27)
j=1
1 − γci j=1
1 − γci j=1
j6=i
Hence it follows from (24)–(27) that when 0 < γ < 1 sufficiently small, we
have
 
 m0
X 
LV (x, i) ≤ γ(1 − γci )|x|γ µi − qbij cj + O(γ)
 
j=1
γ
= γ(1 − γci )|x| (−β + O(γ))
β
≤ − γ(1 − γci ) |x|γ := −W
c (x, i).
2
Note that Ker(W c ) = {0} × M. Therefore we conclude from Theorem 3.7
point x = 0 is asymptotically stable with probability 1. 2
4. Further Remarks
This paper has focused on invariance principles of regime-switching diffu-
sion processes. The results obtained accommodate the stability results and
provide further insight. For subsequent study, we may consider such prob-
lems as the ω-limit sets in the deterministic setup. In lieu of memoryless
setup, if delays are considered in the system, one gets switching diffusions
with delay. Studying invariance for such systems is a worthwhile effort.
Another direction is the investigation of associated problems with Poisson
type of jumps, in which the sample paths are no longer continuous. It will
be interesting to see if any invariance principle can be obtained in that
regards.
Acknowledgments
This research was supported in part by the National Science Foundation
under grant DMS-0907753 and in part by the Air Force Office of Scientific
Research under grant FA9550-10-1-0210.
References
1. G. Barone-Adesi and R. Whaley, Efficient analytic approximation of Ameri-
can option values, J. Finance, 42 (1987), 301–320.
2. G.K. Basak, A. Bisi, and M.K. Ghosh, Stability of Degenerate Diffusions
with State-Dependent Switching, J. Math. Anal. Appl., 240 (1999), 219-248.
3. R.J. Elliott, L. Chan, and T.K. Siu, Option pricing and Esscher transform
under regime switching, Annals of Finance, 1 (2005), 423–432.
4. A. Friedman, Stochastic Differential Equations and Applications, Vols. I and
II, Academic Press, New York, 1975.
5. M.K. Ghosh, A. Arapostathis, and S.I. Marcus, Ergodic control of switching
diffusions, SIAM J. Control Optim., 35 (1997), 1952-1988.
6. J.K. Hale, Sufficient conditions for stability and instability of autonomous
functional differential equations, J. Diff. Equations, 1 (1965), 452–482.
7. J.K. Hale and E.F. Infante, Extended dynamic systems and stability theorem,
Proc. Nat. Acad. Sci., 58 (1967), 405-409.
8. J.D. Hamilton and R. Susmel, Autoregressive conditional heteroskedasticity
and changes in regime. J. Econometrics 64, (1994), 307-333.
9. A.M. Il’in, R.Z. Khasminskii, and G. Yin, Asymptotic expansions of solu-
tions of integro-differential equations for transition densities of singularly
perturbed switching diffusions, J. Math. Anal. Appl., 238 (1999), 516-539.
10. J. Jacod and A.N. Shiryayev, Limit Theorems for Stochastic Processes,
Springer-Verlag, New York, 1980.
11. R.Z. Khasminskii, Stochastic Stability of Differential Equations, Sijthoff and
Noordhoff, Alphen aan den Rijn, Netherlands, 1980.
12. R.Z. Khasminskii, C. Zhu, and G. Yin, Stability of regime-switching diffu-
sions, Stochastic Proc. Appl., 117 (2007), 1037-1051.
13. H.J. Kushner, The concept of invariant set for stochastic dynamical systems
and applications to stochastic stability, in H.F. Karreman, editor, Stochastic
Optimization and Control, John Wiley and Sons, New York, 1968, 47-57.
14. J.P. LaSalle, The extent of asymptotic stability, Proc. Nat. Acad. Sci., 46
(1960), 363–365.
15. X. Mao, Stability of stochastic differential equations with Markovian switch-
ing, Stochastic Proc. Appl., 79 (1999), 45-67.
16. X. Mao, A note on the LaSalle-Type theorems for stochastic differential delay
equations, J. Math. Anal. Appl., 268 (2002), 125–142.
17. X. Mao and C. Yuan, Stochastic Differential Equations with Markovian
Switching, Impreial College Press, London, 2006.
18. M. Pemy, Q. Zhang, and G. Yin, Liquidation of a large block of stock with
regime switching, Math. Finance, 18 (2008), 629-648.
19. A.V. Skorohod, Asymptotic Methods in the Theory of Stochastic Differential
Equations, Amer. Math. Soc., Providence, RI, 1989.
20. H.L. Yang and G. Yin, Ruin probability for a model under Markovian switch-
ing regime, in Probability, Finance and Insurance, World Sci., 206-217, T.L.
Lai, H. Yang, and S.P. Yung Eds., 2004.
21. G. Yin, V. Krishnamurthy, and C. Ion, Regime switching stochastic approxi-
mation algorithms with application to adaptive discrete stochastic optimiza-
tion, SIAM J. Optim., 14 (2004), 1187-1215.
22. G. Yin, X. Mao, C. Yuan, and D. Cao, Approximatio methods for hybrid
diffusion systems with state-dependent switching diffusion processes: Numer-
ical alogorithms and existence and uniqueness of solutions, SIAM Journal on
Mathematical Analysis, 41 (2010), 2335–2352.
23. G. Yin and C. Zhu, Regularity and recurrence of switching diffusions, J. Syst.
Sci. Complexity, 20 (2007), 273-283.
24. G. Yin and C. Zhu, Hybrid Switching Diffusions: Properties and Applications,
Springer, New York, 2010.
25. C. Yuan and X. Mao, Asymptotic stability in distribution of stochastic dif-
ferential equations with Markovian switching, Stochastic Proc. Appl., 103
(2003), 277-291.
26. Q. Zhang, Stock trading: An optimal selling rule, SIAM J. Control Optim.,
40 (2001), 64–87.
27. X.Y. Zhou and G. Yin, Markowitz’s mean-variance portfolio selection with
regime switching: a continuous-time model, SIAM J. Control Optim., 42
(2003), 1466–1482
28. C. Zhu and G. Yin, Asymptotic properties of hybrid diffusion systems, SIAM
J. Control Optim., 46 (2007), 1155-1179.
29. C. Zhu and G. Yin, On strong Feller, recurrence, and weak stabilization of
regime-switching diffusions, SIAM J. Control Optim., 48 (2009), 2003–2031.
30. C. Zhu, G. Yin, and Q.S. Song, Stability of random-switching systems of
differential equations, Quarterly Appl. Math., 67 (2009), 201–220.
May 11, 2011 11:18 WSPC - Proceedings Trim Size: 9in x 6in 04-bens
63
Real Options and Competition
Alain Bensoussan
School of Management
University of Texas at Dallas
The Hong Kong Polytechnic University
Richardson, TX 75083-0688, Hong Kong
Email: alain.bensoussan@utdallas.edu
J. David Diltz
Department of Finance and Real Estate
The University of Texas at Arlington
Arlington, TX 76019, USA
Email: diltz@uta.edu
SingRu (Celine) Hoe

University of Texas at Dallas
Richardson, TX 75083-0688, USA
Email: celinehoe@utdallas.edu
This paper provides a review of our research, Bensoussan et al. Refs. 3,4, which
explore optimal investment policies for irreversible capital investment projects
under uncertainty in monopoly and Stackelberg leader-follower frameworks.
Optimal policies derived for the Stackelberg duopoly case in an incomplete
market are, to our knowledge, heretofore undocumented. We show that com-
petition and market incompleteness have vital impacts on investment decisions.
In our survey, we present models in the case of complete markets and in the case
of incomplete markets with different types of investment payoffs and market
competition frameworks. We discuss model setup, explaining solution tech-
nique and procedure. We provide major results and theorem, and refer proofs
and details to our original papers Refs. 3,4.
Keywords: Real options; optimal stopping; variational inequality; utility based

pricing; differential game.
1. Introduction
This paper provides a review of our research, Bensoussan et al. Refs. 3,4,
which explore optimal investment policies for irreversible capital investment
64 A. Bensoussan et al.
projects under uncertainty in monopoly and Stackelberg leader-follower

frameworks. We present models in the case of complete markets and in
the case of incomplete markets with different types of investment payoffs
and market competition frameworks. We discuss model setup, explaining
solution technique and procedure. We provide major results and theorem.
Among the most significant developments in corporate finance over the
last twenty-five years are arbitrage-free pricing models to valuation and
management of the firm’s fixed assets. Dubbed ”real options” by Myers
(1977), research in this area has since proliferated, and landmark works
are numerous. An essential work in this domain is the book of Dixit and
Pindyck (1994) Ref. 9. Additional references are Constantinides (1978) Ref.
8 applying stochastic optimal control theory to investment decisions, Bren-
nan and Schwartz (1982a, 1982b) Refs. 5,6 who apply stochastic control
to the problem of a regulated public utility, Brennan and Schwartz (1985)
Ref. 7, McDonald and Siegel (1986) Ref. 18, among others.
Two important issues distinguish real options from their financial coun-
terparts. First, competitive strategy is seldom relevant in finance because
financial markets are assumed to be efficient and liquid. No single trader
exerts a measurable impact on the market. However, competitive strategy
is common to real options because capital investment decisions made by one
firm frequently have a significant impact on the cash flows of its competi-
tors (see Dixit and Pindyck (1994, Ch9) Ref. 9 and Grenadier (1996, 1999,
2002) Refs. 10–12). Second, market incompleteness is not a major problem
in financial markets because most financial assets are heavily traded. It is
a critical issue in real options because assets underlying real options are
rarely traded in markets. Financial mathematicians address the market in-
completeness issue by employing strategies such as: (1) minimizing tracking
error, (2) selecting a martingale measure for pricing using minimal martin-
gale or minimal entropy methods, and (3) employing utility functions to
estimate an indifference price. The existing real options literature is quite
limited in addressing competition and incompleteness simultaneously due
to the mathematical complexity.
Ignoring these issues by applying classical real options models may yield
a non-optimal capital investment policy. In our research, Bensoussan et
al. Refs. 3,4, we extend the current financial economics research by simul-
taneously incorporating competitive strategy and market incompleteness.
Our primary results concerning optimal investment policies when competi-
tive strategy and market incompleteness appear simultaneously are, to our
knowledge, new.
Real Options and Competition 65
We consider an irreversible capital investment project under uncertainty

in a Stackelberg leader-follower game, with predetermined leader and fol-
lower. While decisions are made over an infinite horizon, the follower is
forbidden to undertake the investment project until the leader has already
done so. We model uncertainties with respect to two types of investment
payoffs: lump-sum investment payoffs and investment payoffs as a series of
cash flows. The lump-sum payoff is characterized by an externally defined
project value process. The cash flow payoff is generated by either: (1) an
arithmetic Brownian motion cash flow process, or (2) a geometric Brownian
motion cash flow process, where the former captures the possibility of losses
from investment operation.
The firm’s problem is to evaluate the investment project and select the
optimal (stopping) time to invest. We use variational inequalities (V.I.s)
to solve the optimal stopping problems corresponding to investment deci-
sions. V.I. theory was introduced by Stampacchia (1964) and Lions and
Stampacchia (1967) Ref. 17 with applications drawn principally from me-
chanics. Bensoussan and Lions (1982) Ref. 1 were the first to apply V.I.s
to solve stochastic optimal control problems.
The issue of completeness and incompleteness comes in for the valuation
of payoffs. In the case of complete markets, there exists a unique martingale
measure, so the value of future payoffs are simply discounted mathematical
expectations with respect to this unique measure. In the case of incomplete
markets, we adopt the traditional utility maximization approach. The in-
vestor/ entrepreneur maximizes his/her expected discounted exponential
(CARA) utility function with respect to a stopping time for capital invest-
ment, an investment strategy, and/or a consumption rule.
For the lump-sum payoff in the case of incomplete markets, we overcome
the leader’s problem of comparing gains and losses at different times using
certainty equivalence. In this case, we can only formulate and solve leader’s
V.I. in the weak sense. For the case of investment payoffs as a series of
cash flows in incomplete markets, we solve the utility maximization prob-
lem in two steps: (1) we assume that the investment is already in place (i.e.,
the ”post-investment” period), and solve the control problem for a portfo-
lio strategy and a consumption rate; (2) In the second step, we solve the
complete utility maximization problem, choosing an optimal stopping time,
an optimal consumption rule and an optimal portfolio investment strategy
(i.e., the ”pre-investment” period). The idea behind this two-step proce-
dure is that we need the solution of the post-investment period to define
the obstacle function of the V.I. for the pre-investment period. This V.I.
has a nice interpretation as a stochastic differential game, different from

the Stackelberg game.
The follower’s ( and monopolist’s) V.I. has a continuously differentiable
(C 1 ) “obstacle” function (i.e., payoff received at the stopping time). In this
case, the optimal stopping time decision can be obtained by a threshold
approach. The Stackelberg leader’s V.I. contains a non-smooth obstacle
function (C 0 , not C 1 ), and thus we must formulate the V.I. in a weak
sense because the solution of the V.I. has the regularity of the obstacle.
Nevertheless, in most cases, we are able to show a strong solution of the V.I.
exists since the obstacle is not continuously differentiable only at a single
point. Due to the lack of smoothness of the obstacle, the leader’s optimal
strategy (the solution of the V.I.) is a two-interval solution, characterized
by three thresholds.
Except for the case of lump-sum investment payoffs in case of incomplete
markets, we are able to characterize the leader’s optimal investment policies
as three thresholds. To enjoy monopoly rents for a longer period of time,
the leader prefers choosing the lower level of threshold. This understandable
but non-intuitive situation cannot be predicted without the mathematical
theory.
The outline of the paper is as follows. First, we review the valuation and
general features of our research problems and models. We present models
and summarize our research results in the case of complete markets and
in the case of incomplete markets in Sections 3 and 4 respectively. We end
with a conclusion in Section 5.
2. Valuation and General Features of Problems and Models

2.1. Asset Valuation in Complete and Incomplete Markets
The essential property of complete markets is that there exists a uniquely
defined risk-neutral probability measure. In the case of complete markets
we will measure the value of losses and gains by risk-neutral expectation.
For our future reference, the framework is as follows. Consider a probability
space (Ω, F , Q) with a Wiener process W (t). We assume that the market
is characterized by a single asset, S(t), evolving as:
dS(t) = rS(t)dt + σS(t)(λdt + dW (t)), (1)
where r is a risk free interest rate, σ is the volatility and λ is the Sharpe
ratio; they are all constants. The market is complete and there exists a
unique risk-neutral probability measure, obtained as follows. Define:
W
c (t) = W (t) + λt
and the process Z(t) by:
dZ(t) = −λZ(t)dW (t), Z(0) = 1.
The Radon-Nikodym density of the risk neutral probability measure Q

b with
respect to Q is:
dQ
b
|F = Z(t)
dQ t
with Ft = σ(W (s), s ≤ t). Under Q,

b Wc (t) is a standard Wiener process. We
denote E[·]
b as the expectation with respect to the risk neutral probability
measure Q.
b
If capital markets are not complete, the analysis changes. The risk-
neutral probability is not uniquely defined. The classical risk-neutral valu-
ation approach is no longer appropriate. One proposed solution is to intro-
duce a rational utility-maximizing investor who evaluates unhedgeable risk
based on the investor’s risk preferences. Utility-based valuation in stochas-
tic dynamic market environments derives from the famous work of Merton
(1969) Ref. 19. In our incomplete model, we employ the utility-based valua-
tion, assuming that a hypothetical investor/entrepreneur’s risk preferences
may be specified by an exponential (CARA) utility function.
2.2. Market Frameworks

We consider two different market frameworks: a single market player (mo-
nopolist) or two market players (duopolists). Market players’ decisions are
times to enter into the market. In the monopolist’s case, by paying an in-
vestment cost K at the market entry time τ , the firm/entrepreneur expects
to receive the whole operation income.
In the duopolists’ case, we consider a Stackelberg leader-follower game,
with predetermined leader and follower. The follower is forbidden to invest
until the leader has already done so. Each player pays an investment cost
K upon entry. The leader receives the whole operation income prior to the
follower’s entry. Upon follower’s entry, the leader must share market with
follower. In our framework, the portion of share is given.
2.3. Types of Investment Payoffs

We model two types of operating income from the capital investment un-
dertaken. One is in the form of a lump-sum, i.e., the investment project
yields a one time payoff at the time of investment for a given initial invest-
ment K. We model this situation in terms of the investment project value,
governed by an externally determined geometric Brownian motion process.
In this model, the leader has to incorporate into his/her objective function
the surrender value to the follower upon the follower’s optimal entry, so one
has to compare values of events taking place at different times. In the case
of incomplete markets, we work with investor’s utility and there is no iden-
tity relationship for values at different times. We circumvent the problem
by employing equivalence (indifference) considerations.
Alternatively, we consider operating income characterized by cash flows,
i.e., upon paying an initial investment K, the entrepreneur receives a se-
ries of cash flows thereafter. The cash flow evolves as either an arithmetic
Brownian motion process or a geometric Brownian motion process. The
cash flow valuation model circumvents the problem of comparing gains and
losses at different times.
3. The Complete Market Case

3.1. Single Player
3.1.1. Lump-sum Payoffs - External Value Process
The project value defining lump-sum investment payoffs evolves as:
dV (t) = rV (t)dt + ηV (t)(ξdt + dW (t)) (2)

= (r + η(ξ − λ))V (t)dt + ηV (t)dW
c (t), V (0) = v,
where η and ξ are constants.

By paying K at τ , the firm expects to receive the corresponding payoff
(1 − a)Vv (τ ), where a ∈ [0, 1) represents market share by other active par-
ticipants. For the single decision maker case, a is zero. In the Stackelberg
games that follow, a represents the leader’s market share after the follower’s
entry into the market.a
a In the context of the leader-follower model, we assume a > 12 . Otherwise, the leader
will surrender a majority interest in the project, providing a strong disincentive to enter
at all.
The manager’s expected discounted payoff undertaken at time τ is:

b e−rτ (1 − a)Vv (τ ) − K 1τ <∞ .

Jv (τ ) = E (3)
The manager chooses an optimal stopping time to maximize the expected
discounted project value:
F (v) = sup Jv (τ ). (4)
τ ≥0
Solving (4) by V.I., we obtain:


 K v β v ≤ v̂
F (v) = β − 1 v̂ , (5)
 (1 − a)v − K v ≥ v̂
where
s
1 r + η(ξ − λ) 1 r + η(ξ − λ) 2 2r
β= − + − + 2 > 1, (6)
2 η2 2 η2 η
βK
v̂ = (7)
(1 − a)(β − 1)
and we assume ξ < λ.
The optimal stopping rule τ̂ (v) which achieves the supremum in (4) is:
τ̂ (v) = inf{t|Vv (t) ≥ v̂} (8)
with ˆ(v) defined in (7).
3.1.2. Cash Flow Payoffs - Geometric Brownian Motion Cash Flow

Process
The cash flow process Y (t) from investment operation evolves as:
dY (t) = Y (t)(αdt + ςdW (t)) (9)

= Y (t) (α − λς)dt + ςdW
c (t) , Y (0) = y,
where α and ς are constants.

If the firm exploits the investment opportunity by paying cost K, at
time t, the firm expects to receive a continuous cash flow δY (t) per unit
time. The project value at time t, V (t), is:
∞ −r(s−t) δYy (t)
Z

V (t) = δ E
b e Yy (s)ds|Ft = , (10)
t r − (α − λς)
provided r − (α − λς) > 0.
From (10), the manager’s expected discounted payoff from the capital
investment project taken at time τ is:
b e−rτ
δYy (τ )
Jy (τ ) = E − K 1τ <∞ . (11)
r − (α − λς)
The manager’s value function is:
F (y) = sup Jy (τ ). (12)
τ ≥0

 β β−1
δ β−1
y β , y ≤ ŷ



F (y) = β r − (α − λς) K (13)
δy
− K, y ≥ ŷ



r − (α − λς)
where
r
1 α − λς 1 α − λς 2 2r
β= − + − + 2 > 1, (14)
2 ς2 2 ς2 ς
βK r − (α − λς)
ŷ = (15)
β−1 δ
and we assume r − (α − λς) > 0.
The optimal stopping rule τ̂ (y) that achieves the supremum in (12) is:
τ̂ (y) = inf{t|Yy (t) ≥ ŷ} (16)
with ŷ defined in (15).
3.1.3. Cash Flow Payoffs - Arithmetic Brownian Motion Cash Flow

Process
The cash flow process Y (t) from investment operation evolves as:
dY (t) = αdt + ςdW (t) (17)
= (α − λς)dt + ςdW
c (t), Y (0) = y,
where α and ς are constants.

If the firm exploits the investment opportunity by paying cost K, at
time t, the firm expects to receive a continuous cash flow δY (t) per unit
time. The project value at time t, V (t), is:
∞ −r(s−t) Yy (t) α − λς
Z

V (t) = δ E
b e Yy (s)ds|Ft = δ + . (18)
t r r2
From (18), the manager’s expected discounted payoff from the capital
investment project taken at time τ is:

−rτ Yy (τ ) α − λς
Jy (τ ) = E e
b δ( + 2
) − K 1τ <∞ . (19)
r r
The manager’s value function is:
F (y) = sup Jy (τ ). (20)
τ ≥0

δ


 exp{−β(ŷ − y)}, if y ≤ ŷ
F (y) = rβ (21)
 δ y + α − λς − K, if y ≥ ŷ

r r 2
where
r
α − λς α − λς 2 2r
β=− + + 2 > 0, (22)
ς2 ς2 ς
and
1 r α − λς
ŷ = + K− > 0. (23)
β δ r
The optimal stopping rule τ̂ (y) that achieves the supremum in (20) is:
τ̂ (y) = inf{t|Yy (t) ≥ ŷ} (24)
with ŷ defined in (23).
3.2. Two Players: A Stackelberg Leader-Follower Game

There are two players with predetermined leader and follower. The follower
is forbidden to invest until the leader has already done so. The leader enters
at time θ and the follower at time τ ≥ θ where θ and τ are stopping times of
the filtration Ft . Each player invests K upon entry. Upon follower’s optimal
entry, the leader must share market (i.e., project value) with follower. In
our framework, the portion of share is given.

Upon follower’s optimal entry, leader surrenders (1 − a), a ∈ (0, 1) portion
of project value to the follower.
A. Follower
The follower’s problem is similar to the single player’s (the monopolist’s)

since, after the leader enters, the follower makes an optimal stopping deci-
sion as if he/she is a single player. The follower’s solution is the same as
that defined in (8). However, the stopping time τ̂ (v) is the optimal entry if
the follower can enter in the market at time zero. Since the follower enters
after the leader (who starts at θ), for finite θ, the follower will enter at time:
τ̂θ = θ + τ̂ (Vv (θ)).
B. Leader
When the leader enters at time θ < ∞, by paying cost K, he/she receives:
aVv (θ)1Vv (θ)≥v̂ + (Vv (θ) − βF (Vv (θ))1Vv (θ)<v̂ − K,
where F (v), β and v̂ are defined in (5), (6) and (7) respectively. The leader’s
value function is:
b e−rθ Ψ(Vv (θ))1θ<∞ ,

L(v) = sup E (25)
θ≥0
where Ψ(v) = av1v≥v̂ + (v − βF (v))1v<v̂ − K. L(v) must satisfy:

L(v) ≥ 0; L(v) ≥ Ψ(v). (26)
Setting U (v) = L(v) − av + K, (25) can be re-expressed as:
Z θ
−rθ −rs

U (v) = sup E e χ Vv (θ) 1θ<∞ +
b e f Vv (s) ds . (27)
θ≥0 0
This formulation leads to an optimal stopping time problem with an ob-

stacle χ(v) = Ψ(v) − av + K and a running profit f (v) = aη(ξ − λ)v + rK.
Both U (v) and χ(v) are bounded; we have:
0 ≤ χ(v) ≤ K, and 0 ≤ U (v) ≤ K. (28)
The obstacle χ(v) is not C 1 (0, ∞) with a single non-differentiable point at
v̂ and we have to formulate the V.I. in the weak sense.b
Through studying the regularity of the solution U (v) by partial differen-
tial equation techniques, it turns out that the solution U (v) will be smoother
than the obstacle. Therefore, the V.I. will have a strong formulation, and
we can in fact obtain an explicit two-interval solution.
b We refer the formulation and proof to the original work Bensoussan et al. (2009) Ref. 3.
Theorem 3.1. There exist three points 0 < v1 < v2 < v̂ < v3 such that:


 − 21 U 00 (v)η 2 v 2 − (r + η(ξ − λ))vU 0 (v) + rU (v) = aη(ξ − λ)v + rK


 for 0 < v < v and v < v < v
1 2 3


 U (v) = χ(v) for v1 ≤ v ≤ v2

U (v) = 0 for v > v3

(29)
0 0 0 0 0
with matching conditions: U (v1) = χ (v1 ); U (v2 ) = χ (v2 ); U (v3 ) = 0,
where χ(v) = (1 − a)v − βF (v) 1v<v̂ satisfying

1 00 2 2 0
− 2 χ (v)η v − (r + η(ξ − λ))vχ (v) + rχ(v) = −(1 − a)η(ξ − λ)v


if v < v̂

χ(0) = χ(v̂) = (1 − a)v̂ − βK = 0

β−1
(30)
with F (v), β and v̂ defined in (5), (6), and (7) respectively.
We can now define the leader’s optimal stopping rule as:



 inf{t|Vv (t) ≥ v1 } if 0 ≤ v < v1
0 if v1 ≤ v ≤ v2

θ̂(v) = . (31)

 inf{t|Vv (t) ≤ v2 or Vv (t) ≥ v3 } if v2 < v < v3

0 if v ≥ v3

Process
The leader receives a continuous cash flow δ1 Y (t) per unit time prior to
follower’s entry. Upon follower’s optimal entry, each gets a continuous cash
flow δ2 Y (t) per unit time.
A. Follower
The follower’s solution is the same as that defined in (16) but replacing δ
by δ2 . Since the follower enters after the leader (who starts at θ < ∞), the
follower will enter at time: τ̂θ = θ + τ̂ (Yy (θ)).
B. Leader

δ2 Yy (θ) Yy (θ) β
−K + + (δ1 − δ2 ) − F (Yy (θ)) 1Yy (θ)<ŷ
r − α + λς r − α + λς δ2
where F (y), β and ŷ are is defined in (13), (14), and (15) respectively. The
leader’s value function is:

L(y) = sup E{exp(−rθ)Ψ
b Yy (θ) 1θ<∞ } (32)
θ≥0

δ2 y y β
where Ψ(y) = −K + r−α+λς + (δ1 − δ2 ) r−α+λς − δ2 F (y) 1y<ŷ . L(y) must
satisfy:
L(y) ≥ 0; L(y) ≥ Ψ(y). (33)

δ2 y
Setting U (y) = L(y) − r−α+λς + K, (32) can be re-expressed as:
Z θ
U (y) = sup E e−rθ χ Yy (θ) 1θ<∞ + −rs

b e f Yy (s) ds . (34)
θ≥0 0
This formulation leads to an optimal stopping time problem with an obsta-

δ2 y
cle χ(y) = Ψ(y) − + K and a running profit f (y) = −δ2 y + rK.
r − α + λς
Both U (y) and χ(y) are bounded; we have:
δ1 − δ2
0 ≤ U (y) ≤ K max(1, ),
δ2
δ1 − δ2 δ1 − δ2
and 0 ≤ χ(y) ≤ K ≤ K max(1, ). (35)
δ2 δ2
The obstacle χ(y) is not C 1 (0, ∞) with a single non-differentiable point at

ŷ and we thus have to formulate the V.I. in the weak sense.c
Applying the regularity study as in Section 3.2.1, it turns out that the so-
lution U (y) will be smoother than the obstacle. Therefore, the V.I. will have
a strong formulation and we can obtain an explicit two-interval solution.
Theorem 3.2. There exist three points 0 < y1 < y2 < ŷ < y3 = ȳ such
c We refer the formulation and proofs to the original work Bensoussan et al. (2009) Ref. 4.
that:


 − 21 U 00 (y)ς 2 y 2 − (α − ςλ)yU 0 (y) + rU = −δ2 y + rk


 for 0 < y < y and y < y < y
1 2 3


 U (y) = χ(y) for y1 ≤ y ≤ y2

U (y) = 0 for y ≤ y3

with matching conditions: U 0 (y1 ) = χ0 (y1 ); U 0 (y2 ) = χ0 (y2 ); U 0 (y3 ) = 0,

y
where χ(y) = (δ1 − δ2 ) r−α+λς − δβ2 F (y) 1y<ŷ satisfying
(
− 21 χ00 (y)ς 2 y 2 − (α − λς)yχ0 (y) + rχ(y) = (δ1 − δ2 )y, for y < ŷ
(36)
χ(0) = χ(ŷ) = 0
with F (y), β and ŷ defined in (13), (14), and (15) respectively. The function
U vanishes for y sufficiently large.
The leader’s optimal stopping rule is defined in the the same way as the
lump-sum payoff case (cf.(31)).

Process
The leader receives a continuous cash flow δ1 Y (t) per unit time prior to
follower’s entry. Upon follower’s optimal entry, each gets a continuous cash
flow δ2 Y (t) per unit time.
A. Follower
The follower’s solution is the same as that defined in (24) but replacing δ
by δ2 . Since the follower enters after the leader (who starts at θ < ∞), the
follower will enter at time: τ̂θ = θ + τ̂ (Yy (θ)).
B. Leader
Yy (θ) α − λς Yy (θ) α − λς
−K + δ2 + 2
+ (δ1 − δ2 )1Yy (θ)<ŷ +
r r r r2
ŷ α − λς −β(ŷ−Yy (θ))

− + e ,
r r2
where β and ŷ are defined in (22) and (23) respectively. The leader’s value
function is:
b e−rθ Ψ(Yy (θ))1θ<∞

L(y) = sup E (37)
θ≥0
δ2 yr
+ α−λς + (δ1 − δ2 )1y<ŷ yr + α−λς 1

where Ψ(y) = −K + r 2 r2 − rβ +
K
−β(ŷ−y)
δ2 e . L(y) must satisfy:
L(y) ≥ 0; L(y) ≥ Ψ(y). (38)

Ψ(y) is unbounded as y → ±∞. The obstacle, Ψ(y) is not C 1 (−∞, ∞) with
the only non-differentiable point at ŷ, so we have to formulate the V.I. in
the weak sense.d
Again, applying the regularity study, it turns out that the solution L(y)
will be smoother than the obstacle. Therefore, the V.I. will have a strong
formulation and we can obtain an explicit two-interval solution.
Theorem 3.3. Define:

δ2 δ2
u(y) = L(y) − ( y + 2 (α − ςλ)) + K
r r
and
δ2 δ2
m(y) = Ψ(y) − ( y + 2 (α − ςλ)) + K
r r
with Ψ(y) satisfying:
(
− 12 ς 2 Ψ00 (y) − (α − ςλ)Ψ0 (y) + rΨ = δ1 y − rK, y < ŷ
δ2 δ2 δ2
(39)
Ψ(ŷ) = r ŷ + r 2 (α − ςλ) − K = rβ
and m(y) satisfying

1
− ς 2 m00 (y) − (α − ςλ)m(y) + rm(y) = (δ1 − δ2 )y, y < ŷ. (40)
2
There exists a triple y1 < y2 < ŷ < y3 such that:

1 2 00 0
− 2 ς u (y) − (α − ςλ)u (y) + ru = −δ2 y + rK, y < y1 and y2 < y < y3


u(y) = m(y), y1 ≤ y ≤ y2

u(y) = 0, y ≥ y

3
with matching conditions: u0 (y1 ) = m0 (y1 ); u0 (y2 ) = m0 (y2 ); u0 (y3 ) = 0.

The function u vanishes for y sufficiently large.
d We refer the formulation and proofs to the original work Bensoussan et al. (2009) Ref. 3.
lump-sum payoff case (cf.(31)).
4. The Incomplete Market Case

4.1. Utility-Based Pricing
The asset S representing the market still evolves as (1). However, the
project value defining lump-sum investment payoffs now evolves according
to:
p
dV (t) = rV (t)dt + ηV (t) ξdt + ρdW (t) + 1 − ρ2 dW 0 (t) ,

(41)
the cash flow investment payoffs by geometric Brownian motion process

evolve as:
p
dY (t) = Y (t) αdt + ς(ρdW (t) + 1 − ρ2 dW 0 (t)) , Y (0) = y,

(42)
and the cash flow investment payoffs by arithmetic Brownian motion pro-
cess evolve as:
p
dY (t) = αdt + ς ρdW (t) + 1 − ρ2 dW 0 (t) , Y (0) = y

(43)
where W (t) and W 0 (t) are independent Wiener processes. The parameter
|ρ| < 1 is the correlation coefficient between market uncertainty and in-
vestment payoff uncertainty. The market asset S can only span a portion
of the investment payoff risk driven by the Wiener process W (t), leaving
the remaining risk driven by W 0 (t) unhedgeable. The unique equivalent
martingale measure is undefined, so the risk-neutral pricing is no longer
appropriate, and an alternative must be developed in this framework.
We adopt the utility-based pricing approach where the risk averse in-
vestor’s preferences are characterized by an exponential utility function
given by:
1
U (x) = − e−γx (44)
γ
where γ is his/her risk aversion parameter, γ > 0.
The valuation formulation is: given initial wealth, the investor consid-
ers utility maximization by choosing an optimal stopping time to invest
the project together with a portfolio investment-consumption strategy. We
will formulate the investor’s specific utility maximization problem in each
corresponding scenarios.
4.2. Single Player

The rational investor maximizes his/her expected utility of wealth. Given
initial wealth, x, the risk averse investor optimizes his/her portfolio by
dynamically choosing allocations in the market asset S and the riskless
bond. The investor’s wealth X, evolves as:

dX(t) = rX(t)dt + X(t)π(t)σ λdt + dW (t) , (45)
where π(t) is the portion of wealth invested in asset S. We use the dis-
counted value Ve (t) = V (t)e−rt and X(t)
e = X(t)e−rt where:
p
dVe (t) = Ve (t)η ξdt + ρdW (t) + 1 − ρ2 dW 0 (t) , Ve (0) = v

(46)

dX(t) = X(t)π(t)σ λdt + dW (t) , X(0) = x.
e e e (47)
Let Ft = σ(W (s), W 0 (s); s ≤ t). We consider stopping times τ with respect
to Ft . At τ , the individual invests and receives X ex (τ ) + (Vev (τ ) − K)+e . To
a pair (π(·), τ ), we associate with the objective function:

λ2
Jx,v (π(·), τ ) = E e 2 τ U Xex (τ ) + Vev (τ ) − K + .

(48)
We assume τ is finite a.s.. The function (48) is well defined but the value
−∞ is possible.
We define the associated value function as:
F (x, v) = sup Jx,v (π(·), τ ). (49)

π(·),τ

( 1
U (x) χ(v) 1−ρ2 , if 0 < v < v̂
F (x, v) = (50)
U (x)e−γ (1−a)v−K , if v ≥ v̂
where
v β
χ(v) = 1 − (1 − e−$ ) , (51)
v̂
2(ξ − ρλ)
β =1− > 1, (52)
η
e This formulation is similar to the specification that the investment cost grows at the
risk-free rate.
v̂ solves
$
(1 − a)v̂ = K + (53)
γ(1 − ρ2 )
with $ the unique positive solution of
$ Kγ(1 − ρ2 )
e$ − =1+ , (54)
β β
and we assume ξ − ρλ < 0.
We next define the optimal stopping rule which achieves supremum in
(49):
ex (t) + (1 − a)Vev (t) − K +

τ̂ (x, v) = inf t ≥ 0|F X
ex (t), Vev (t) = U X
= inf{t ≥ 0|Vev (t) is outside (0, v̂)} = τ̂ (v) (55)
and τ̂ (v) < ∞ a.s..

Process
The rational investor maximizes his/her expected utility of consumption.f
Given the initial wealth, x, the risk averse investor optimizes his/her port-
folio by dynamically choosing allocations in the market asset S, the riskless
bond, and the consumption rate, C. The investor’s wealth X, evolves as:



 dX(t) = π(t)X(t)σ(λdt + dW (t)) + rX(t)dt − C(t)dt, t < τ

X(τ ) = X(τ − 0) − K



dX(t) = π(t)X(t)σ(λdt + dW (t)) + rX(t)dt − C(t)dt + δY (t)dt, t > τ ,
 p
dY (t) = Y (t) αdt + ς(ρdW (t) + 1 − ρ2 dW 0 (t))






X(0) = x, Y (0) = y
(56)
where π(t) is the proportion of wealth invested in asset S, C(t) is the
consumption rate, and τ is a stopping time to undertake the investment,
chosen optimally by the investor. The wealth process is discontinuous at τ .
From (56), we observe that the wealth process has two possible evolution
regimes, so we introduce:
dX 0 (t) = π(t)X 0 (t)σ(λdt + dW (t)) + rX 0 (t)dt − C(t)dt, (57)
f We allow for negative consumption.

and
dX 1 (t) = π(t)X 1 (t)σ(λdt + dW (t)) + rX 1 (t)dt − C(t)dt + δY (t)dt, (58)
where X 0 is the wealth process X before the stopping time τ and X 1 is the
wealth process X after the stopping time τ .
A. Post-Investment Utility Maximization
We begin with a control problem relative to the process X 1 (t). Introduce

1
the set of admissible controls, Ux,y = {C(·), π(·)}, satisfying:
( RT 2
E 0 π(t)X 1 (t) dt < ∞ ∀ T
RT 2 , (59)
E 0 C(t) dt < ∞ ∀ T
τN ↑ ∞ a.s. as N ↑ ∞, where τN = inf{t|X 1 (t) ≤ −N }, (60)

and a transversality condition:

X 1 (T )+f (Y (T ))
e−µT E[e−rγ ] → 0, as T → ∞ (61)
where µ is the discount rate, f (y) is a positive function which will be
made precise later (cf.(69),(68)) and f (y) has linear growth with f (0) = 0
(cf.(67)). To a pair of C(·), π(·) , we introduce the objective function:
Z ∞
e−µt U C(t) dt].

J C(·) = E[ (62)
0
This function is well defined, but it may take the value −∞. We define the
value function:
F 1 (x, y) = sup J C(·)). (63)
1
{π(·),C(·)}∈Ux,y
First, we note that if y = 0, then Y (t) = 0, ∀ t. The problem reduces to

the classical investment-consumption problem with the solution given by:
2
1 µ + λ2
F (x) = − exp{−rγx + 1 − }. (64)
rγ r
We have:
F 1 (x, 0) = F (x). (65)
We look for a solution of (63) as follows:
λ
1 µ+
F 1 (x, y) = − exp{−rγ(x + f (y)) + 1 − 2
} (66)
rγ r
in which, by (65), implies:

f (0) = 0. (67)
By (66), the Bellman equation that the value function F 1 (x, y) (63) must
satisfy reduces to:
1 2 2 00 1
y ς f + (α − λςρ)yf 0 − rγy 2 ς 2 (1 − ρ2 )f 02 − rf + δy = 0. (68)
2 2
Proposition 4.1. The function,
Z ∞
1
e−rt δYy (t) + v 2 (t) dt

f (y) = inf E (69)
{v(·)∈Uy } 0 2
with
( p
dY (t) = Y (t) α − λςρ + ς rγ(1 − ρ2 )v(t) dt + Y (t)ςdW (t), Y (0) = y
R ∞ −rt
Uy = {v(·)|E[ 0 e v(t)2 dt] < ∞, e−rT E[Yy (T )] → 0 as T → ∞}
is the unique function in C 2 (0, ∞), solution of (68), (67) in the interval
[0, y + M ]g , such that f (y) ↑ ∞ as y ↑ ∞. Moreover, the function f 0 (y) is
bounded.
Note that the function f (y) defined in (69) is the function entering into the
transversality condition in (61).
Theorem 4.1. The function F 1 (x, y) given by (66) coincides with the value
function given by (63).
B. Pre-Investment Utility Maximization
We now turn to the problem of optimal stopping with the obstacle defined
by F 1 (x, y). Before the stopping time τ , the wealth process X is governed
by X 0 (t) (57) and the cash flow process evolves as (42).
0
Introduce the set of admissible controls, Ux,y = {π(.), C(.), τ }, satisfy-
ing:
( RT 2
E[ 0 π(t)X 0 (t) dt] < ∞, ∀ T
RT , (70)
E[ 0 C(t)2 dt] < ∞, ∀ T
τ ∧ θ0 < ∞ a.s. where θ0 = inf{t|Y (t) = 0}, (71)
δ
2
gM
− r + α − λςρ
is defined as: M = . Note that can be arbitrarily small.
2ς 2 r 2 γ(1 − ρ2 )
and τ ∗ = lim ↑ τN ≥ τ ∧ θ0 a.s. where τN = inf{t|X 0 (t) ≤ −N }. (72)
If θ0 ≤ τ, the investment never takes place and the investor receives
F X 0 (θ0 ) (cf.(64)). If τ < θ0 , the investor will receive F 1 (X 0 (τ ) −
K, Y (τ ))(cf.(66)) at the stopping time τ . Therefore, the objective function
is:
Z τ ∧θ0
e−µt U C(t) dt + F 1 X 0 (τ ) − K, Y (τ )

Jx,y C(·), π(·), τ = E
0

−µτ 0 0
−µθ0
×e 1τ <θ0 + F X (θ ) e 1θ0 ≤τ
(73)
and we define the associated value function:

F (x, y) = sup Jx,y C(·), π(·), τ . (74)
0
{π(·),C(·),τ }∈Uxy
We look for a solution of (74) in the form:

2
1 µ + λ2
F (x, y) = − exp − rγ x + g(y) + 1 − . (75)
rγ r
The V.I. that the value function F (x, y) (74) must satisfy reduces to:
 1 2 2 00 0 1 2 2 2 02

 2 y ς g + g y(α − λςρ) − 2 y ς rγ(1 − ρ )g − rg ≤ 0
 g(y)
 ≥ f (y) − K

g(y) − f (y) + K 12 y 2 ς 2 g 00 + g 0 y(α − λςρ) − 21 y 2 ς 2 rγ(1 − ρ2 )g 02 − rg

=0





g(0) = 0.
(76)
In (76), the nonlinear operator is connected to a minimization problem and
the inequalities are connected to a maximization problem for a stopping
time. g(y) can be interpreted as the value function of a differential game.
Define: u(y) = g(y) − f (y) + K, yielding:

− 12 y 2 ς 2 u00 − yu0 α − λςρ − yf 0 ς 2 rγ(1 − ρ2 ) + 12 y 2 ς 2 rγ(1 − ρ2 )u02 + ru







 ≥ −δy + rK
u≥ 0

u − 12 y 2 ς 2 u00 − yu0 α − λςρ − yf 0 ς 2 rγ(1 − ρ2 ) + 12 y 2 ς 2 rγ(1 − ρ2 )u 02







 + ru + δy − rK = 0
u(0) = K.

(77)
We study (77) by the threshold approach. For fixed ŷ, we solve the Dirichlet
problem:
 1 2 2 00 0 0 2 2
1 2 2 2 02
 − 2 y ς u − yu α − λςρ − yf ς rγ(1 − ρ ) + 2 y ς rγ(1 − ρ )u + ru
= −δy + rK, 0 < y < ŷ
u(0) = K, u(ŷ) = 0.

(78)
Equation (78) is a Bellman equation of the following control problem with
the controlled diffusion:
p
dY = Y α − λςρ − Y f 0 (Y )ς 2 rγ(1 − ρ2 ) + ς rγ(1 − ρ2 )v dt + Y ςdW

Y (0) = y, 0 < y < ŷ

(79)
and the value function

Z θy v(·)
1

e−rt − δYy (t) + rK + v 2 (t) dt + K × e−rθy v(·)

u(y) = inf E
v(·) 0 2

×1
Yy θy (v(·)) =0
(80)

where θy v(·) = inf{t|Y (t) is outside (0, ŷ)}, and it is a.s. finite.
Theorem 4.2. There exists a unique value ŷ such that
Kr
u0 (ŷ) = 0, ŷ ≥ . (81)
δ
The value function u(y)(cf.80) extended by 0 beyond ŷ is the unique solution
of the V.I. (77). It is C 1 and piecewise C 2 .
Hence, there exists a unique ŷ such that:


 − 21 y 2 ς 2 g 00 − g 0 y(α − λςρ) + 21 y 2 ς 2 rγ(1 − ρ2 )g 02 + rg = 0, y < ŷ


g(y) = f (y) − K, y ≥ ŷ
.


 g 0 (ŷ) = f 0 (ŷ)

g(0) = 0

(82)
Note that g(y) ≥ 0 since u(y) ≥ −f (y) + K.
Theorem 4.3. The function F (x, y) defined by (75) coincides with the
value function given by (74).
We next define the optimal stopping rule as: τ̂ (y) = inf{t|Yy (t) ≥ ŷ}, where
ŷ is the unique value defined by the V.I. (77) and (81)(the smooth matching
point).

Process
As in Section 4.2.2, the rational investor maximizes his/her expected utility
of consumption by dynamically choosing allocations in the market asset S,
the riskless bond, and the consumption rate, C. Given the initial wealth,
x, his/her wealth X evolves the same as (56).h And again, before the stop-
ping time τ , his/her wealth process X corresponds to X 0 (57), and after
the stopping time τ , it corresponds to X 1 (58).
A. Post-Investment Utility Maximization
As in Section 4.2.2, we begin with a control problem relative to the pro-

cess X 1 (t). Introduce the set of admissible controls, Ux,y
1
= {C(·), π(·)},
satisfying:
( RT 2
E 0 π(t)X 1 (t) dt < ∞ ∀ T
RT 2 , (83)
E 0 C(t) dt < ∞ ∀ T
τN ↑ ∞ a.s. as N ↑ ∞, where τN = inf{t|rX 1 (t) + δY (t) ≤ −N }, (84)

and a transversality condition:

rX 1 (T )+δY (T )
e−µT E[e−γ ] → 0, as T → ∞. (85)
To a pair of (C(·), π(·)), we introduce the objective function:
Z ∞
e−µt U C(t) dt].

J C(·) = E[ (86)
0
This function may take the value −∞. We define the associated value func-
tion as:
F 1 (x, y) =

sup J C(·) . (87)
1
{π(·),C(·)}∈Ux,y
Theorem 4.4.
2
1 δ µ + λ2
F 1 (x, y) = −

exp − rγ(x + y) + 1 −
rγ r r
δγ 1
α − λςρ − ς 2 δγ(1 − ρ2 )

−
r 2
(88)
is the solution of the value function (87).
p
h The cash flow process now evolves as: dY (t) = αdt + ς(ρdW (t) + 1 − ρ2 dW 0 (t)).
B. Pre-Investment Utility Maximization
We now solve the optimal stopping problem with the obstacle defined by
the value function F 1 (x, y). Before the stopping time τ , the wealth pro-
cess X 0 (t) evolves as (57) and the cash flow process Y (t) evolves as (43).
0
Introduce the set of admissible controls Ux,y = {π(·), C(·), τ } satisfying:
2
( RT
E[ 0 π(t)X 0 (t) dt] < ∞, ∀ T

RT , (89)
E[ 0 C(t)2 dt] < ∞, ∀ T
τ < ∞ a.s., and τ ∗ = lim ↑ τN ≥ τ a.s. whereτN = inf{t|X 0 (t) ≤ −N }.

(90)
At time τ , the investor undertakes the investment and receives F 1 (X 0 (τ ) −
K, Y (τ ))(cf.(88)). Therefore, the objective function is:
Z τ
e−µt U C(t) dt + F 1 X 0 (τ ) − K, Y (τ ) e−µτ

Jx,y C(·), π(·), τ = E
0
(91)
and we define the value function:

F (x, y) = sup Jx,y C(·), π(·), τ . (92)
0
{π(·),C(·),τ }∈Uxy
We look for a solution of the form:

2
1 µ + λ2
F (x, y) = − exp − rγ x + g(y) + 1 − . (93)
rγ r
2
1 µ+ λ
Recall F 1 (x, y) = − rγ exp − rγ x + f (y) + 1 − r 2 .i The V.I. that the

value function F (x, y) (92) must satisfy reduces to:

 1 2 00 1
 2 ς g + g 0 (α − λςρ) − 2 ς 2 rγ(1 − ρ2 )g 02 − rg ≤ 0
g(y) ≥ f (y) − K .
1 2 00 0 1 2 2 02

g(y) − f (y) + K 2 ς g + g (α − λςρ) − 2 ς rγ(1 − ρ )g − rg = 0

(94)
Considering u(y) = g(y) − f (y) + K, the V.I. (94) becomes:
 1 2 00
− 2 ς u − u0 α − λςρ − ς 2 δγ(1 − ρ2 ) + 12 ς 2 rγ(1 − ρ2 )u02 + ru





 ≥ −δy + rK
u≥ 0 . (95)
 u − 12 ς 2 u00 − u0 α − λςρ − ς 2 δγ(1 − ρ2 ) + 12 ς 2 rγ(1 − ρ2 )u02




+ ru + δy − rK = 0
δy δ
i f (y) α − λςρ − 21 ς 2 ργ(1 − ρ2 ) .

= r
+ r2
We study (95) by the threshold approach. For ŷ fixed, we solve the Dirichlet
problem:
 1 2 00 0 2 2
1 2 2 02
 − 2 ς u − u α − λςρ − ς δγ(1 − ρ ) + 2 ς rγ(1 − ρ )u + ru
= −δy + rK, y < ŷ
u(ŷ) = 0

(96)
and we require a linear growth for y → −∞.
Theorem 4.5. For each ŷ, there exists a unique solution of (96) with the
estimate:
(−δ ŷ + rK)−

δ 1
− ≤ u(y) ≤ − (y − ŷ)+ − δ ŷ + rK
r r r
+

δ
− α − λςρ − γδς 2 (1 − ρ2 ) ,
r
for y < ŷ. (97)
The solution is C 2 in (−∞, ŷ). There exists a unique ŷ such that:
rK
u0 (ŷ) = 0, ŷ ≥ . (98)
δ
The corresponding solution of (96) extended by 0 beyond ŷ is the unique
solution of the V.I. (95). It is C 1 and piecewise C 2 .
Hence, there exists a unique ŷ such that:


1 2 00 0 1 2 2 02
− 2 ς g − g (α − λςρ) + 2 ς rγ(1 − ρ )g + rg = 0, y < ŷ


g(y) = rδ y − K + rδ2 α − λςρ − 21 ς 2 δγ(1 − ρ2 ) , y ≥ ŷ . (99)

g 0 (ŷ) = δ

r
We have also: g(y) → 0 as y → −∞, and g(y) ≥ 0.
Theorem 4.6. The function F (x, y) given by (93) coincides with the value
function (92).
We next define the optimal stopping rule, which achieves supremum in

(92), as: τ̂ (y) = inf{t|Yy (t) ≥ ŷ}, where ŷ is the unique value defined by
the V.I. (95) and (98).
4.3. Two Players: A Stackelberg Leader-Follower Game

The setting is the same as Section 3.2.

A. Follower
The follower’s strategy is identical to that described in the Section 4.2.1. We

have the follower’s value function as defined in (50). The optimal stopping
strategy for the follower is:
τ̂ (v) = inf{t ≥ 0|Vev (t) is outside (0, v̂)},
where v̂ is defined in (53). The stopping time τ̂ (v) is the optimal entry if
the follower can enter in the market at time zero. Since the follower enters
after the leader (who
starts at θ), for finite θ, the follower will enter at time:
τ̂θ = θ + τ̂ Vev (θ) .
B. Leader
As the case of a single player, the leader dynamically optimizes his/her

investment portfolio and solves his/her utility maximization problem with
joint decisions of stopping times and portfolio investment strategies. When
the leader enters at time θ < ∞,
(1) If Ve (θ) > v̂, the follower enters immediately and the leader gets X(θ)
e +
aVe (θ) − K.
(2) If Ve (θ) < v̂, then τ̂ Ve (θ) > 0, the leader gets immediately X(θ)e +

(Ve (θ) − K), but he/she surrenders at time θ + τ̂ Ve (θ) , a percentage
(1 − a) of project value to the follower.
A difficulty occurs in the latter scenario. At time θ, the leader must deter-
mine the value surrendered to the follower, taking into account the follower’s
optimal entry at τ̂θ . However, we work with investor’s utility and there is
no identity relationship for values at different times. We will circumvent
this problem converting the surrender value by employing equivalence (in-
difference) considerations.
Let X(t),
e Ve (t) be the wealth of the follower and the value process
governed by (47) and (46). Suppose θ = 0 and v > v̂, then the follower
receives (1 − a)v at time 0 and he/she values this operation by U x + (1 −

a)v). If θ = 0 and 0 < v < v̂, then the follower receives (1 − a)Ve τ̂ (v) at
λ2
time τ̂ (v); the corresponding value is: e 2 τ̂ (v) U X

ex τ̂ (v) +(1−a)Vev τ̂ (v) .
Since the follower can dynamically optimize his/her portfolio between time
0 and τ̂ (v), the value is in fact:

λ2
H(x, v) = max E e 2 τ̂ (v) U X

ex τ̂ (v) + (1 − a)Vev τ̂ (v) (100)
π(·)

and we have the boundary conditions: H(x, v̂) = U x + (1 − a)v̂ and
H(x, 0) = U (x).
Proposition 4.2. The solution of (100) is:
H(x, v) = U (x)Λ(v) (101)

1 − $+Kγ(1−ρ2 )
β 1−e

1−ρ2
where: Λ(v) = 1 + Bv and B = − v̂ β
, with β, v̂, $
defined in (52), (53) and (54) respectively.
We next define the follower’s indifference value H e (v) by:

e
H(x + H e (v), 0) = H(x, v) = U (x)e−γH (v)
; (102)
then:
 −γH e (v)
e = Λ(v), 0 < v < v̂
H e (v) ≤ (1 − a)v, if v ≤ v̂ . (103)
 e
H (v) = (1 − a)v, v ≥ v̂
H e (v) is the value that makes the follower indifferent between receiving it
at time 0 and losing the right to undertake investment at time τ̂ (v), or to
exert the right to undertake the investment. The leader must deduct this
amount in addition to K at time 0 if he/she decides to invest at time 0.
We now formulate completely the leader’s problem. For each pair of
(π(·), θ), the leader’s objective function is :

λ2 +
Jx,v (π(·), θ) = E e 2 θ U X(θ)
e + Ve (θ) − K − H e (Ve (θ)) (104)
and we define associated value function:

L(x, v) = sup Jx,v π(·), θ . (105)
π(·),θ
We impose θ < ∞ a.s. We look for a solution of the form:

1
L(x, v) = U (x)L(v) = U (x)Σ(v) 1−ρ2 with 0 ≤ L(v) ≤ 1. (106)
We obtain the V.I.:


0 1 00
Σ (ξ − λρ) + 2 vηΣ ≥ 0


 +
2 e
Σ(v) ≤ e−γ(1−ρ ) v−K−H (v)



 +
2
Σ (ξ − λρ) + 12 vηΣ00 Σ(v) − e−γ(1−ρ )
0 v−K−H e (v)
=0 . (107)





 0 ≤ Σ(v) ≤ 1


Σ(0) = 1
+
2 e
The obstacle, ψ(v) = e−γ(1−ρ ) v−K−H (v) is continuous but not C 1 ,
and its derivative is discontinuous at v o (K < v o < v̂) and v̂. Therefore, we
must consider (107) in a weak sense. Nevertheless, (107) is the V.I. of the
following optimal stopping problem with the state equation:

dV (t) = V (t)η (ξ − λρ)dt + dW (t) , V (0) = v (108)
and the value function:

Σ(v) = inf Jv (τΣ ) with Jv (τΣ ) = E ψ Vv (τΣ ) (109)
τΣ
and τΣ must be a stopping time which is a.s. finite. To formulate (109) in a

weak form sense, we introduce the Sobolev space H%1 (0, ∞) with the scalar
0
R∞ (v)φ̃0 (v)
product:((φ, φ̃))% = (φ, φ̃)% + 0 v 2 φ(1+v 2 )% dv. We define the bilinear form:
∞
1 − v 2 (% − 1) φ̃(v)
Z
φ0 (v) − η(ξ − λρ) + η 2

b(φ, φ̃) = dv
0 1 − v2 (1 + v 2 )%
1 ∞ φ0 (v)φ̃0 (v)v 2 η 2
Z
+ dv.
2 0 (1 + v 2 )%
The V.I. corresponding to (109) is:
b(Σ, Σ
e − Σ) ≥ 0, ∀Σ
e ∈ K, Σ ∈ K (110)
where K = {φ ∈ H%1 (0, ∞)|0 ≤ φ(v) ≤ ψ(v), ∀ v} is the convex subset of

H%1 (0, ∞).
Theorem 4.7. Assume ξ − λρ < 0. Then there exists one and only one
solution of (110). It coincides with the value function (109).
Because of the presence of the term H e (v), we cannot have a general

result as it is the case of complete markets.

Process
A. Follower
We have the follower’s value function as defined by (75) with g(y) satisfying
(76) and δ replaced by δ2 . The optimal stopping strategy for the follower
is:
τ̂ (y) = inf{t|Y (t) ≥ ŷ}
where ŷ is a fixed number defined by the V.I. (77),and (81)(the smooth

matching point). We must take δ = δ2 and ŷ > rKδ2 . Since the follower can
(who starts at θ < ∞), the follower will enter at

enter only after the leader
time: τ̂θ = θ + τ̂ Yy (θ) .
B. Leader
By investing K, the leader expects to receive a continuous cash flow δ1 Y (t)

per unit time prior to the follower’s entry, and δ2 Y (t) per unit time after-
wards.
B.1 Post-Investment Utility Maximization
As in Section 4.2.2, we solve the leader’s utility maximization problem in

two steps, beginning with the control problem assuming the capital invest-
ment project has been undertaken.
Suppose θ = 0, the leader’s wealth is x, and the cash flow y > 0; then
the leader’s wealth becomes immediately x − K since he/she must pay the
fixed cost of entry, K. The leader must share the market upon follower’s
entry at τ̂ (y). Thus, for a generic initial wealth x, the leader’s wealth evolves
as:

dX 1 (t) = π(t)X 1 (t)σ λdt + dW (t) + rX 1 (t)dt + δ1 Y (t)dt − C(t)dt,




for 0 < t < τ̂ (y)





X 1 (0) = x

dX 2 (t) = π(t)X 2 (t)σ λdt + dW (t) + rX 2 (t)dt + δ2 Y (t)dt − C(t)dt,





for t > τ̂ (y)





 2 1

X τ̂ (y) = X τ̂ (y) .
(111)
If θ = 0 and y > ŷ, the follower enters immediately, and the leader’s
problem is identical to the follower’s, i.e., (63) with δ = δ2 . So, we consider
the function:
2
µ+ λ
1

2
L2 (x, y) = − e−rγ x+f (y) +1− r , (112)
rγ
where f is the solution of (68), (67) in the interval [0, y + M]j with δ being
replaced by δ2 .
If θ = 0 and y < ŷ, the leader’s problem is described as follows. The
wealth process is described by X 1 in (111) and the cash flow process follows
1
(42). Introduce the set of admissible controls, Ux,y = {C(·), π(·)}, satisfying:
( RT 2
E 0 π(t)X 1 (t) dt < ∞, ∀ T < ∞
RT , (113)
E 0 C(t)2 dt < ∞, ∀ T < ∞

and
τ ∗ = lim ↑ τN ≥ τ̂ (y) ∧ θ0 a.s. where

τN = inf{t|X 1 (t) ≤ −N } and τ̂ (y) = inf{t|Yy (t) ≥ ŷ}. (114)
Recall θ0 = inf{t|Yy (t) = 0}. If θ0 < τ̂ (y), the follower never in-
vests, and the leader’s value function at time θ0 corresponds to
F (X 1 (θ0 ))(cf.(64)). Ifτ̂ (y) < θ0 , the leader’s value function corresponds to
L2 (X 1 τ̂ (y)), Y (τ̂ (y)) , at the follower’s entry time, τ̂ (y). Thus, to a pair
of (C(·), π(·)), we associate the objective function:
0
τ̂ (y)∧θ −µt
Z
0
e U (C(t))dt + F X 1 (θ0 ) e−µθ 1θ0 ≤τ̂ (y)

J C(·), π(·) = E
0
+ L2 X 1 τ̂ (y) , Y τ̂ (y) e−µτ̂ (y) 1τ̂ (y)<θ0 .

(115)
We consider the value function:
L1 (x, y) =

sup J C(·), π(·) . (116)
1
{π(·),C(·)}∈Ux,y
We study the Bellman equation associated with (116) for 0 < y < ŷ. We
look for a solution of the form:
2
µ+ λ
1

2
L1 (x, y) = − e−rγ x+g(y) +1− r , (117)
rγ
j See footnote 4.1 for the definition of M .

and we define:
L1 (x, y) = L2 (x, y), if y > ŷ,
2
where L (x, y) is defined in (112). The Bellman equation that the value
function L1 (x, y) (116) must satisfy reduces to:
(
1 2 2 00 0 1 2 2 2 2 2 02
2 y ς g + (α − λςρ)yg − 2 r γ y ς (1 − ρ )g − rg + δ1 y = 0, 0 < y < ŷ
g(0) = 0, g(ŷ) = f (ŷ).
(118)
k
Where f (y) is the solution of (68), (67) in the interval [0, y + M ] with δ
being replaced by δ2 . We extend g(y) by f (y) for y > ŷ.
Similar to the study of (68), g(y) may be interpreted as a function of a
control problem, and there exists a unique solution which is C 2 (0, ŷ).
Theorem 4.8. The function L1 (x, y) defined by (117) coincides with the
value function given in (116).
B.2 Pre-Investment Utility Maximization
We now turn to the leader’s optimal stopping problem (i.e., choice of θ)

with obstacle defined by L1 (x, y). Before the stopping time θ, the leader’s
wealth evolves according to (57) and the cash flow evolves as (42). Introduce
0
the set of admissible controls, Ux,y = {C(·), π(·), θ}, satisfying
2
( RT
E 0 π(t)X 0 (t) dt < ∞, ∀ T < ∞

RT , (119)
E 0 C(t)2 dt < ∞, ∀ T < ∞

θ ∧ θ0 < ∞ a.s. where θ0 = inf{t|Yy (t) = 0}, (120)
and τ ∗ = lim ↑ τN ≥ θ ∧ θ0 a.s. where τN = inf{t|X 0 (t) ≤ −N }. (121)

If θ0 ≤ θ, the leader never takes the investment and receives
F (X 0 (θ0 ))(cf.(64)). If θ < θ0 , the leader receives L1 (X 0 (θ) − K, Y (θ))
(cf.(117)) at θ.
Therefore, the objective function is:
0
θ∧θ
Z
U C(t) e−µt dt + L1 X 0 (θ) − K, Y (θ) ×

Jx,y C(·), π(·), θ = E
0
0
e−µθ 1θ<θ0 + F X 0 (θ0 ) e−µθ 1θ0 ≤θ

(122)
k See footnote 4.1 for the definition of M .


L(x, y) = sup Jx,y C(·), π(·), θ . (123)
0
{π(·),C(·),θ}∈Ux,y
We look for a solution as follows:

2
µ+ λ
1

2
L(x, y) = − e−rγ x+h(y) +1− r . (124)
rγ
The V.I. that the value function L(x, y) (123) must satisfy reduces to:

1 1
 2 y 2 ς 2 h00 + h0 y(α − λςρ) − 2 y 2 ς 2 rγ(1 − ρ2 )h02 − rh ≤ 0



h(y) ≥ g(y) − K



h(y) − g(y) + K 21 y 2 ς 2 h00 + h0 y(α − λςρ) − 21 y 2 ς 2 rγ(1 − ρ2 )h02 − rh





 =0


h(0) = 0.
(125)
The obstacle g(y) − K is C 0 but not C 1 . Thus the V.I.(125) must be inter-
preted in a weak sense. Considering: u(y) = h(y) − f (y) + K, we yield:

− 21 y 2 ς 2 u00 − y α − λςρ − yς 2 rγ(1 − ρ2 )f 0 u0 + 12 ς 2 rγy 2 (1 − ρ2 )u02 + ru




≥ −δ2 y + rK






u ≥ m
(u − m) − 21 y 2 ς 2 u00 − y α − λςρ − yς 2 rγ(1 − ρ2 )f 0 u0




+ 12 ς 2 rγy 2 (1 − ρ2 )u02 + ru + δ2 y − rK = 0






u(0) = K

(126)
where m = g(y) − f (y) is the solution of:

− 21 y 2 ς 2 m00 − y α − λςρ − yς 2 rγ(1 − ρ2 )f 0 m0 + 21 ς 2 rγy 2 (1 − ρ2 )m02





 + ry = (δ − δ )y, 0 < y < ŷ
1 2


 m(0) = m(ŷ) = 0

m(y) = 0, y > ŷ.

(127)
We next show that u(y) is more appropriately the value function of a
stochastic differential game.
r−α+λςρ
Theorem 4.9. We assume γς 2 (1−ρ2 ) > δ1 ŷ. There exists a unique u(y) ∈
1 2
C (0, ∞), piecewise C , solution of (126). This function vanishes for y
sufficiently large. Moreover, it is the value function given

u(y) = inf sup Jy v(·), θ = sup inf Jy v(·), θ (128)
v(·) θ θ v(·)
with
 p
2 2 0

 dY (t) = Y (t) α − λςρ − ς rγ(1 − ρ )Y (t)f (Y (t)) + v(t)ς rγ(1 − ρ 2 ) dt



+ ςY (t)dW (t)




Y (0) = y
 R θ0 ∧θ 1 2
−rt 0
e dt + Ke−rθ 1θ0 <θ




 J y v(·), θ = E 0 − δ 2 Yy (t) + rK + 2 v (t)
+ m Yy (θ) e−rθ 1θ<θ0


where θ0 = inf{t|Yy (t) = 0}, and m is defined previously, the solution of

(127) and extended by 0 for y > ŷ.
There exists a saddle point v̂(·), θ̂ such that:

Jy v̂(·), θ̂ = inf sup Jy v(·), θ = sup inf Jy v(·), θ
v(·) θ θ v(·)
and we have 0 ≤ u(y) ≤ K + m̄, where m̄ = sup m(y) ≤ δ1 −δ r ŷ with

2
m(y) = g(y) − f (y), solution of (127) and extended by 0 for y > ŷ.
The solution of (126) is characterized by two intervals.
Theorem 4.10. There exists a unique triple y1 , y2 , y3 with 0 < y1 < y2 <
ŷ < y3 such that
1 1 2
− y 2 ς 2 u00 − y α − λςρ − yς 2 rγ(1 − ρ2 )f 0 u0 + ς rγy 2 (1 − ρ2 )u02 + ru

2 2
= −δ2 y + rK,
0 < y < y1 and y2 < y < y3
(129)
with the smooth pasting conditions:
0 0

 u(y1 ) = m(y1 ), u (y1 ) = m (y1 )
u(y2 ) = m(y2 ), u0 (y2 ) = m0 (y2 ) , (130)
u(y3 ) = 0, u0 (y3 ) = 0

where m(y) = g(y)−f (y), the solution of (127) and extended by 0 for y > ŷ.
Theorem 4.11. The function L(x, y) defined by (124) coincides with the
value function (122).
lump-sum payoff case in case of complete markets (cf.(31)).

Process
A. Follower
The follower’s value function as defined by (93) with g(y) satisfying (94),
where ŷ is the unique value defined by the V.I. (95) and (98). We take
δ = δ2 . The optimal stopping stopping strategy for the follower is:
τ̂ (y) = inf{t|Yy (t) ≥ ŷ},
where ŷ is the unique value defined by the V.I. (95) and (98). Note again
that we must take δ = δ2 and thus ŷ ≥ rK δ2 . Since the follower enters
after the leader (who starts at θ < ∞), the follower will enter at time:
τ̂θ = θ + τ̂ Yy (θ) .
B. Leader
By paying cost K, the leader expects to receive a continuous cash flow

δ1 Y (t) per unit time prior to the follower’s entry, and δ2 Y (t) per unit time
after the follower’s entry.
B.1 Post-Investment Utility Maximization
Suppose θ = 0, his/her wealth is x, and the cash flow y > 0. The wealth
becomes immediately x − K since he/she has to pay the fixed cost of entry.
The follower will enter at τ̂ (y). The leader’s wealth evolves as (111).
If θ = 0, and y ≥ ŷ, the follower enters immediately and the leader’s
problem is exactly the same as the follower’s, i.e., (87) with δ = δ2 . So we
consider the function:
2
µ+ λ
1

2
L (x, y) = − e−rγ
2 x+f (y) +1− r (131)
rγ
δ2 y δ2 1
+ 2 α − λςρ − ς 2 δ2 γ(1 − ρ2 ) .

with: f (y) = (132)
r r 2
If θ = 0 and y < ŷ, the leader’s problem is described as follows. The

leader’s wealth evolves according to (58) with δ being replaced by δ1 and the
cash flow process evolves as (43). Introduce the set of admissible controls
1
Ux,y = {C(·), π(·)} satisfying:
( RT 2
E 0 π(t)X 1 (t) dt < ∞, ∀ T < ∞
RT , (133)
E 0 C(t)2 dt < ∞, ∀ T < ∞

and τ ∗ = lim ↑ τN ≥ τ̂ (y) a.s. where τN = inf{t|X 1 (t) ≤ −N }. (134)

At τ̂ (y), the leader gets L2 X 1 τ̂ (y) , Y τ̂ (y)

(cf.(131)). To a pair
(C(·), π(·)), we associate the objective function:
τ̂ (y) −µt
Z
e C(t)dt + L2 X 1 τ̂ (y) , Y τ̂ (y) e−µτ̂ (y)

J C(·), π(·) = E
0
(135)
where τ̂ (y) < ∞ a.s. We consider the value function:
L1 (x, y) =

sup J C(·), π(·) . (136)
1
{π(·),C(·)}∈Ux,y
We look for a solution:

2
µ+ λ
1

2
L (x, y) = − e−rγ
1 x+g(y) +1− r . (137)
rγ
The V.I. that the value function L1 (x, y) (136) must satisfy reduces to:
(
− 12 ς 2 g 00 − (α − ςρ)g 0 + 12 r2 γ 2 ς 2 (1 − ρ2 )g 02 + rg = δ1 y, y < ŷ
g(ŷ) = f (ŷ), g has linear growth at −∞.
(138)
Considering the difference m = g − f , (138) becomes:

1 2 00 0 2 2
1 2 2 02
− 2 ς m − m α − λςρ − ς δ2 γ(1 − ρ ) + 2 ς rγ(1 − ρ )m + rm


= (δ1 − δ2 )y

m(ŷ) = 0,

m has linear growth at y → −∞.
(139)
Proposition 4.3. There exists one and only one solution of (139) in the
interval:
δ1 − δ2 δ1 − δ2
(y − y ∗ ) ≤ m(y) ≤ y − y0 + (y0 − ŷ)eβ(y−ŷ) , for y < ŷ,

r r
(140)
where β > 0 is the the solution of: − 12 ς 2 β 2 −β α−λςρ−γδς 2 (1−ρ2 ) +r = 0,

2
(1−ρ2 ) δ22 γς 2 (1−ρ2 )
and y0 = − α−λςρ−δ2rγς with f (y0 + r2 ) = 0, and we take:
∗ γς 2 (1−ρ2 )(δ1 −δ2 )
y = max ŷ, y0 + 2r .
Theorem 4.12. The function L1 (x, y) defined by (137) coincides with the
value function given in (136).
B.2 Pre-Investment Utility Maximization
We now turn to the leader’s optimal stopping problem (i.e., choice of θ).
Before the stopping time θ, the leader’s wealth evolves according to (57)
0
and cash flow evolve as (43). Introduce the set of admissible controls, Ux,y =
{C(·), π(·), θ}, satisfying:
( RT 2
E 0 π(t)X 0 (t) dt < ∞, ∀ T < ∞
RT , (141)
E 0 C(t)2 dt < ∞, ∀ T < ∞

θ < ∞ a.s., and τ ∗ = lim ↑ τN ≥ θ a.s. where τN = inf{t|X 0 (t) ≤ −N }.

(142)
At time θ, the leader invests and receives L1 (X 0 (θ) − K, Y (θ))(cf.((137));

the leader’s objective function is:
Z θ
U C(t) e−µt dt + L1 X 0 (θ) − K, Y (θ) e−µθ

Jx,y C(·), π(·), θ = E
0
(143)

L(x, y) = sup Jx,y C(·), π(·), θ . (144)
0
{π(·),C(·),θ}∈Ux,y
We look for a solution of the form:

2
µ+ λ
1

2
L(x, y) = − e−rγ x+h(y) +1− r . (145)
rγ
The V.I. that the value function L(x, y) (144) must satisfy reduces to:

1 2 00 0 1 2 2 02
 2 ς h + (α − λςρ)h − 2 ς rγ(1 − ρ )h − rh ≤ 0


h(y) ≥ g(y) − K .

 h(y) − g(y) + K 1 ς 2 h00 + (α − λςρ)h0 − 1 ς 2 rγ(1 − ρ2 )h02 − rh = 0

2 2
(146)
The obstacle g(y) − K is C 0 but not C 1 , so that the V.I.(146) must be

interpreted in a weak sense. Considering: u(y) = h(y) − f (y) + K, we yield:


− 21 ς 2 u00 − α − λςρ − ς 2 δ2 γ(1 − ρ2 ) u0 + 12 ς 2 rγ(1 − ρ2 )u02 + ru





≥ −δ2 y + rK




u≥m .

− m) − 12 ς 2 u00 − α − λςρ − ς 2 δ2 γ(1 − ρ2 ) u0 + 12 ς 2 rγ(1 − ρ2 )u02



 (u


 + ru + δ y − rK = 0 2
(147)
The function m(y) = g(y) − f (y) is defined by (139). To exclude the case
that the leader and the follower have the same strategy, we will consider:
m is not always negative. (148)
As in the geometric Brownain motion cash flow case, we next show that
u(y) is more appropriately the value function of a stochastic differential
game.
Theorem 4.13. Assume (148). There exists a unique u ∈ C 1 (−∞, ∞),

piecewise C 2 solution of (147). This function vanishes for y sufficiently
large. It is the value function:

u(y) = inf sup Jy v(·), θ = sup inf Jy v(·), θ = Jy v̂(·), θ̂ , (149)
v(·) θ θ v(·)
where
 Rθ 1 2
−rt
Jy v(·), θ = E 0 − δ2 Yy (t) + rK + 2 v (t) e dt + m Yy (θ)


× e−rθ 1θ<∞ .


−rT
v(·) ∈ U = {lim sup
T →∞ − EYy (T )e

y = 0}
(150)
δ2
Moreover, u(y) + r y is bounded for y → −∞, and u ≥ 0.
The solution to (147) is characterized by two intervals.
Theorem 4.14. Assume (148). There exists a unique triple y1 , y2 , y3

with y1 < y2 < ŷ < y3 such that


− 21 ς 2 u00 (y) − α − λςρ − ς 2 δ2 γ(1 − ρ2 ) u0 (y) + 21 ς 2 rγ(1 − ρ2 )u02 (y)




+ ru(y) = −δ2 y + rK for y < y1 and y2 < y < y3






u(y) = m(y) for y ≤ y ≤ y
1 2
0 0
.


 u (y1 ) = m (y1 )

u0 (y2 ) = m0 (y2 )





u(y) = 0 for y ≥ y3

(151)
We can take y3 = ȳ = kŷ with k sufficiently large.
Theorem 4.15. Assume (148). Then the function L(x, y) defined by (145)
is the value function (144).
The leader’s optimal stopping rule can be defined in the same way as
the lump-sum payoff case in case of complete markets (cf.(31)).
5. Conclusion
From our research results, we caution that, for managers, naively assum-
ing market completeness and ignoring the potential strategic interactions
from competitors will likely lead to non-optimal investment decisions. The
leader’s three threshold solution, which is understandable but non-intuitive,
cannot be predicted without the mathematical theory.
References
1. Bensoussan, A. and J.L. Lions, Applications of Variational Inequalities in
Stochastic Control, Elsevier North-Holland, 1978.
2. Bensoussan, A., 2008, ”Real Options”, Handbook of Mathematical Modelling
and Numerical Methods in Finance, A. Bensoussan and Q. Zhang(Eds.),
15(1), Elsevier(December), 531-572.
3. Bensoussan, A., J. D. Diltz and S. Hoe, 2009, ”Real Options Games in Com-
plete and Incomplete Markets with Several Decision Makers,” forthcoming,
SIAM Journal of Financial Mathematics.
4. Bensoussan, A., J. D. Diltz and S. Hoe, 2009, ”Real Options in a Stackelberg
Game with a Stochaastic Demand Process,” to be published.
5. Brennan, M. and E. Schwartz, 1982, Regulation and Corporate Investment
Policy, Journal of Finance, 37(2), 289-300.
6. Brennan, M. and E. Schwartz, 1982, Consistent Regulatory Policy Under
Uncertainty, Bell Journal of Economics, 13(2), 506-524.
7. Brennan, M. and E. Schwartz, 1985, Evaluating Natural Resource Invest-
ments, Journal of Business, 58(2), 135-157.
8. Constantinides, George M., 1978, ”Market Risk Adjustment in Project Val-

uation”, Journal of Finance, 33(2), 603-616.
9. Dixit, A. and R.S. Pindyck, Investment Under Uncertainty, Princeton Uni-
versity Press, 1994.
10. Grenadier, S., 1996, ”The Strategic Exercise Of Options: Development Cas-
cades And Overbuilding In Real Estate Markets,” Journal of Finance, 51(5),
1653-1679.
11. Grenadier, S., 1999, ”Information Revelation Through Option Exercise,” Re-
view of Financial Studies, 12(1), 95-129.
12. Grenadier, S., 2002, ”Option Exercise Games: An Application To The Equi-
librium Investment Strategies Of Firms,” Review of Financial Studies, 15(3),
691-721.
13. Henderson, V., 2008, ”Valuing the Option to Invest in an Incomplete Mar-
ket,” Mathematics and Financial Economics, 1( 2), 103-128.
14. Henderson, V. and D. Hobson, 2002, Real Options with Constant Relative
Risk Aversion, Journal of Economic Dynamics and Control, 27(2), 329-355.
15. Huisman, K.J.M., Technology Investment: A Game Theoretical Real Options
Approach, Kluwer Academic Pbulishing, 2001.
16. Karatzas, I. and S. Shreve, Methods of Mathematical Finance, Springer, 1998.
17. Lions, J.L., and G. Stampacchia, 1967, Variational Inequalities, Comm. P.
Appl. Math., XX, 493-519.
18. McDonald, R. and D. Siegel, 1986, The Value of Waiting to Invest, Quarterly
Journal of Economics, 101(4), 708-728.
19. Merton, R.C., 1969, ”Lifetime Portfolio Selection Under Uncertainty: The
Continuous Time Case”, Review of Economics and Statistics, 51, 247-257.
20. Miao, J. and N. Wang, 2007, ”Investment, Consumption and Hedging under
Incomplete Markets”, Journal of Financial Economics, 86, 608-642.
21. Musiela, M. and T. Zariphopoulou, 2001, ”Pricing and Risk Management of
Derivatives Written on Non-Traded Assets,” Working Paper.
22. Oberman, A. and T. Zariphopoulou, 2003, ”Pricing Early Exercise Contracts
in Incomplete Markets,” Computational Management Science, 75-107.
May 11, 2011 11:25 WSPC - Proceedings Trim Size: 9in x 6in 05-ross
101
Finding Expectations of Monotone Functions of Binary Random

Variables by Simulation, with Applications to Reliability,
Finance, and Round Robin Tournaments
Mark Brown
City College, CUNY, New York, NY, USA
Email: cybergarf@aol.com
Erol A. Peköz
Boston University
595 Commonwealth Avenue
Boston, MA 02215, USA
Email: pekoz@bu.edu
Sheldon M. Ross
Department of Industrial and Systems Engineering
University of Southern California
Los Angeles, CA 90089, USA
Email: smross@usc.edu
We study the quantity E[φ(X1 , . . . , Xn )] when X1 , . . . , Xn are independent

Bernoulli
Pn random variables, and φ is a nondecreasing function. With T =
i=1 Xi , we note that E[φ(X1 , . . . , Xn )|T ] is a nondecreasing function of T ,
and show how it can be efficiently estimated by a simulation study that strat-
ifies on T . Our results are applied to static and dynamic reliability systems,
the pricing of derivatives related to basket default swaps, and to round robin
tournaments.
1. Introduction
Let X , . . . , Xn be independent Bernoulli random variables, and let T =
Pn 1
i=1 Xi be their sum. Also, suppose that φ is a nondecreasing function
of the vector X = (X1 , . . . , Xn ). In Section 2 we show how we can effi-
ciently estimate E[φ(X)] by a simulation approach that stratifies on T.
In Section 3, we note that E[φ(X)|T ] is a nondecreasing function of T ,
and show how this result can sometimes be used to modify our simulation
102 M. Brown, Erol A. Peköz and S. M. Ross
estimates. In Section 4 we apply our results to the classical reliability prob-

lem of finding the probability that a system composed of n independent
binary components will function. In Section 5 we consider a dynamic ver-
sion of the preceding model, in which each component works for a random
length of time and then fails. Depending on the structure of the system,
the component failures eventually cause the system to fail. Supposing that
there is a cost incurred if the system fails before some specified time t∗ ,
with the cost depending on the time of failure, we present a stratification
approach to estimate the expected cost incurred. In Section 6 we apply our
results to the pricing of derivatives related to basket default swaps, where
our methods are similar to, but an improvement on, a recently proposed
simulation approach (see2 ). In Section 7 we apply our results to estimating
win probabilities in round robin tournaments.
2. Estimating E[φ(X1 , . . . , Xn )] by Simulation

Suppose that X = (X1 , . . . , Xn ) is a vector of independent Bernoulli ran-
dom variables with E[Xi ] = pi , and that we are interested in using simula-
Pn
tion to estimate E[φ(X)]. Our approach is to stratify on T = i=1 Xi . To
do so we need to first determine the probability mass function of T , and
then show how to simulate X conditional on T = i, for i = 0, . . . , n.
To start, we note that if
Pj (i) = P (Xj + . . . + Xn = i)
then, analogous to Example 3.22 of,6 Pj (i) satisfy
Pj (i) = pj Pj+1 (i − 1) + (1 − pj )Pj+1 (i) (1)
Pn (1) = pn , Pn (0) = 1 − pn , Pn (i) = 0 , i 6= 0, 1.
Starting with the preceding expression for Pn (i), the equations (1) are easily
solved by a recursion that first solves for Pn−1, then Pn−2 , and so on.
To generate (X1 , . . . , Xn ) conditional on T = i, generate in sequence
• X1 given T = i
• X2 given T = i, X1
• X3 given T = i, X1 , X2
and so on. To generate Xj , given both that T = i and the values of

Finding Expectations of Monotone Functions 103
X1 , . . . , Xj−1 , use that
n
X j−1
X
P (Xj = 1|T = i, X1 , . . . , Xj−1 ) = P (Xj = 1| Xk = i − Xk )
k=j k=1
Pj−1
pj Pj+1 (i − 1 − k=1 Xk )
= Pj−1 . (2)
Pj (i − k=1 Xk )
Suppose one is planning on doing r simulation runs. One way to employ

the stratification approach is to do rP1 (i) runs conditional on T = i for each
i = 0, . . . , n. Call the preceding the “standard proportional stratification”
approach, Because T and φ(X) are both increasing functions of the vector
of independent random variables X it follows that they are positively cor-
related and thus the standard proportional stratification will result in an
estimator with smaller variance than the estimator based on r non-stratified
runs (see7 ).
However, a more efficient approach than using standard proportional
stratification is to first do a preliminary study by simulating φ(X) condi-
tional on T = i enough times so as to get a rough estimate of
s2i ≡ Var(φ(X)|T = i) , i = 0, . . . , n.
Then, if you are planning to do r simulation runs, do r PsisPk1P(i)

1 (k)
of these
k
runs conditional on T = i. (That this results in the minimal variance of the
final estimator see.7 ) Letting φi be the average of the runs done conditional
Pn Pn
on T = i, estimate E[φ(X)] = i=0 E[φ(X)|T = i]P1 (i) by i=0 φi P1 (i).
Remark. Whether using the standard or the more efficient procedure, fur-
ther variance reduction can be obtained by using antithetic variables. That
is, in generating (X1 , . . . , Xn ) conditional on T = i, first generate random
numbers U1 , . . . , Un , and then generate the value of Xj , given both that
T = i and the values of X1 , . . . , Xj−1 , by letting it equal 1 if Uj is less than
the right side of (2) In the next generation of (X1 , . . . , Xn ) conditional on
T = i, we can utilize the same set of random numbers, but this time sub-
tracting each from 1. (That is, we use the random numbers 1−U1, . . . , 1−Un
in the next run.) Because φ is a monotone function the reuse of the random
number set will lead to a variance reduction when compared with using a
new independent set of random numbers (see7 for a proof).
Example 1. Let Xi , i = 1, . . . , 20, be independent Bernoulli random vari-

ables with common mean P (Xi = 1) = 1/2, and let
X20
φ(x1 , . . . , x20 ) = ( ixi )2 .
i=1
The standard deviation of the standard Monte Carlo (commonly referred to

as the raw) simulation estimator of E[φ(X1 , . . . , Xn )] based on 104 runs is
57, whereas the standard deviation of the stratified estimator based on the
same number of runs is 26, a variance reduction by a factor of approximately
4. However, noting that the lower indexed Xi have a much smaller effect
on the value of φ(X1 , . . . , Xn )] than do the higher ones, it seems reasonable
P20
that rather than stratifying on i=1 Xi it would be better to stratify on,
P20 P20
say, i=10 Xi . A further simulation indicated that stratifying on i=10 Xi
reduced the standard deviation (again based on 104 runs) down to 8, a
variance reduction over the standard Monte Carlo estimator by a factor of
about 50.
3. Monotonicity of the Conditional Expectation Given T

Not only, as noted in the preceding section, are T and φ(X1 , . . . , Xn ) posi-
tively correlated but, even stronger, is that the conditional distribution of
φ(X) given that T = k is stochastically increasing in k. That is, we have
the following result.
Theorem 3.1. If X1 , . . . , Xn are independent Bernoulli random variables,

and φ a nondecreasing function, then E[φ(X1 , . . . , Xn )|T = k] is a non-
decreasing function of k.
Theorem 3.1 is a special case of a general result of 3 which states that the
preceding is true whenever the Xi are independent and have logconcave
densities or mass functions. As the proof of 3 is rather involved, we now
present a proof that is not only elementary but also in the spirit of this
paper. We being with a lemma.
Lemma 3.1. The conditional distribution of Xn given T is stochastically

increasing in T . That is, P (Xn = 1|T = k) is a nondecreasing function of
k.
Proof. The proof makes use of the log concavity result that P (T =
k)/P (T = k − 1) is nonincreasing in k. (For a proof, see.5 ) Let pn =
Pn−1
P (Xn = 1) = 1 − qn . Also, let br = P ( i=1 Xi = r). Then, we have
pn bk−1
P (Xn = 1|T = k) =
P (T = k)
pn bk−1
=
pn bk−1 + qn bk
pn
=
pn + qn bk /bk−1
and the result follows since bk /bk−1 is nonincreasing by log concavity.
We now prove the theorem.

Pr
Proof of Theorem 3.1. Let Sr = i=1 Xi . Also, for fixed k > 0,
let Y1 , ..., Yn be distributed as X1 , ..., Xn conditional on Sn = k; and let
Z1 , ..., Zn be distributed as X1 , ..., Xn conditional on Sn = k − 1. We now
show how we can generate random vectors Y1 , ..., Yn and Z1 , ..., Zn that
are distributed according to the preceding and are such that Yi ≥ Zi , i =
1, . . . , n. To do so, first note that by the Lemma 3.1
P (Yn = 1) ≥ P (Zn = 1).
We generate the random vectors as follows:
KEY STEP Generate a random number U . Then,

if U ≤ P (Zn = 1), set Zn = 1, else set it equal to 0,
if U ≤ P (Yn = 1), set Yn = 1, else set it equal to 0.

There are now 3 cases:
Case 1: Zn = Yn = 1
Given the scenario of this case, Y1 , ..., Yn−1 is distributed as X1 , .., , Xn−1
conditional on Sn−1 = k − 1; and Z1 , ..., Zn−1 is distributed as X1 , .., , Xn−1
conditional on Sn−1 = k − 2. Consequently, by Lemma 3.1
P (Yn−1 = 1|Yn = 1) ≥ P (Zn−1 = 1|Zn = 1).
Thus, we can repeat KEY STEP to generate Yn−1 and Zn−1 so that
Yn−1 ≥ Zn−1 .
Case 2: Zn = Yn = 0
Given the scenario of this case, Y1 , ..., Yn−1 is distributed as X1 , .., , Xn−1
conditional on Sn−1 = k; and Z1 , ..., Zn−1 is distributed as X1 , .., , Xn−1

conditional on Sn−1 = k − 1. Consequently, by Lemma 3.1
P (Yn−1 = 1|Yn = 0) ≥ P (Zn−1 = 1|Zn = 0).
Thus, we can repeat KEY STEP to generate Yn−1 and Zn−1 so that
Yn−1 ≥ Zn−1 .
Case 3. Zn = 0, Yn = 1
Given the scenario of this case, the random vectors Y1 , ..., Yn−1 and
Z1 , ..., Zn−1 have the same joint distribution. Thus, we can generate them
so that Yi = Zi , i = 1, . . . , n − 1.
The preceding shows how to generate the vectors so that Yi ≥ Zi , i =

1, . . . , n. By the monotonicity of φ this implies that φ(Y1 , . . . , Yn ) ≥
φ(Z1 , . . . , Zn ). Consequently,
E[φ(X1 , . . . , Xn )|T = k]= E[φ(Y1 , . . . , Yn )]
≥ E[φ(Z1 , . . . , Zn )]
= E[φ(X1 , . . . , Xn )|T = k − 1].
Remark. While Theorem 1 might seem quite intuitive, it does depend on

the Xi being Bernoulli random variables. For instance, suppose that X1
and X2 both put all their mass on the values 0, 2, 3. Then, with
φ(x1 , x2 ) = x21 + x22
we would have that
E[φ(X1 , X2 )|T = 3] > E[φ(X1 , X2 )|T = 4].
Suppose now that our simulation of the preceding section resulted in

the estimate φi of E[φ(X)|T = i] for i = 0, . . . , n. If it results that φi is
not nondecreasing in i, then we can modify these estimates by using the
ideas of isotonic regression. Isotonic regression takes preliminary estimates
e1 , . . . , en of unknown quantities that are known to be nondecreasing, and
obtains final estimates a1 , . . . , an by solving the minimization problem
n
X
min (ei − ai )2 ,
a1 ≤...≤an
i=1
which can generally be solved in time linear in n (see1 for details).
The following corollary will be used in the sequel.
Corollary 3.1. Suppose that the random vectors (Xi , Wi ), i = 1, . . . , n are

independent, where Xi , i = 1, . . . , n are Bernoulli random variables, and
where
Wi |Xi = 1 ≥st Wi |Xi = 0
where by the preceding we mean that P (Wi ≥ y|Xi = 1) ≥ P (Wi ≥ y|Xi =

Pn
0), for all y. Let T = i=1 Xi . Then, for any nondecreasing function h,
E[h(W1 , . . . , Wn )|T = k] is a nondecreasing function of k.
Proof. Let g(X) = E[h(W)|X] where W = (W1 , . . . , Wn ) and X =

(X1 , . . . , Xn ). Because the random vectors are independent and Wi given
Xi = 1 is stochastically larger than Wi given Xi = 0, it follows that g(X)
is a nondecreasing function of X. Hence, by Theorem 3.1, it follows that
E[g(X)|T = k] is a nondecreasing function of k. The result now follows
because
E[h(W)|T = k] = E[E[h(W)|T = k, X]|T = k]

= E[E[h(W)|X]|T = k]
= E[g(X)|T = k].
4. The Classical Reliability Model

Consider an n component system in which each component is either working
or failed, and suppose that there exists a nondecreasing binary function
φ such φ(x1 , . . . , xn ) is 1 if the system works when xi is the indicator
variable for whether component i is working, i = 1, . . . , n. The function
φ is called the structure function, and to rule out trivialities we assume
that φ(0, 0, . . . , 0) = 0 and φ(1, 1, . . . , 1) = 1. Now consider the problem
of determining E[φ(X1 , . . . , Xn )] when the Xi are independent Bernoulli
Pn
random variables with pi = E[Xi ], i = 1, . . . , n. With T = i=1 Xi , we can
use the approach of the preceding sections to estimate E[φ(X1 , . . . , Xn )] by
doing a simulation that stratifies on T .
Call any set of components having the property that the system nec-
essarily works when all of these components are working a path set. If no
proper subset of a path set is itself a path set, call the path set a minimal
path set. Let m1 denote the size of the smallest minimal path set, and let
m2 be the size of the largest minimal path set. Because the system works
whenever T > m2 and does not work when T < m1 , we have that
m2
X
E[φ(X1 , . . . , Xn )] = E[φ(X1 , . . . , Xn )|T = i]P (T = i) + P (T > m2 ).
i=m1
Consequently, we need only estimate the quantities E[φ(X1 , . . . , Xn )|T = i]

for m1 ≤ i ≤ m2 , and this can be done using the approach of Section 2. If r
rP (T =i)
simulation runs are planned then one can either do P (m 1 ≤T ≤m2 )
runs condi-
tional on T = i, m1 ≤ i ≤ m2 or, better, do an initial small size simulation
to estimate the conditional variances Var(φ(X1 , . . . , Xn )|T = i), m1 ≤ i ≤
m2 and then do a larger simulation in which the number of runs done con-
ditional on T = i is proportional to P (T = i) times the square root of
the estimate of the Var(φ(X1 , . . . , Xn )|T = i). If the resulting estimates of
E[φ(X1 , . . . , Xn )|T = i] are not monotone in i, then the estimates can be
modified by an isotonic regression.
5. A Dynamic Reliability Model

Again suppose that φ is a structure function for an n component system.
Suppose that each of the n components is initially working, and that com-
ponent i works for random time Wi , i = 1, . . . , n. In addition, suppose that
W1 , . . . , Wn are independent, with Wi having distribution function Fi . If L
is the amount of time that the system itself works, then
L = max min Wj
i=1,...,s j∈Mi
where M1 , . . . , Ms are the minimal path sets for the structure function φ.
Suppose that a cost C(t) is incurred if the lifetime of the system is t,
where, for some specified time t∗ ,
C(t) = h((t∗ − t)+ )
where h is a nondecreasing function having h(0) = 0. In other works, there

is no cost if system life exceeds t∗ and a cost h(s) if the system fails at time
t∗ − s, s < t∗ . We are interested in using simulation to estimate E[C(L)].
Let Xi equal 1 if component i is still working at time t∗ and let it
equal 0 otherwise. Then E[φ(X)] is the probability that the system life
Pn
exceeds t∗ . Let T = ∗
i=1 Xi . With pi = 1 − Fi (t ), i = 1, . . . , n, let
Pj (i), i, j = 1, . . . , n be the solution of (1).
We propose to estimate E[C(L)] by stratifying on T . To begin, let m

be the size of the largest minimal path set. Because the system cannot be
failed if there are more than m working components, we have
m
X
E[C(L)] = E[C(L)|T = i]P1 (i).
i=0
To simulate C(L) conditional on T = i , i = 0, . . . , m, first use the

method given in Section 2 to generate X1 , . . . , Xn conditional on T = i. If
φ(X1 , . . . , Xn ) = 1 then take 0 as the estimate of E[C(L)|T = i] from that
run. If φ(X1 , . . . , Xn ) = 0, then for any j for which Xj = 0, generate Wj
according to the distribution
Fj∗ (t) = P (Wj ≤ t|Wj ≤ t∗ ) = Fj (t)/Fj (t∗ ) , 0 ≤t ≤ t∗.
For j such that Xj = 1, set Wj = t∗ . With L being the lifetime of the

system according to the preceding values of Wj , take C(L) as the estimate
of E[C(L)|T = i] from that run. Of course the number of runs to do con-
ditional on T = i should be determined by a preliminary small simulation
study to estimate the quantities Var(C(L)|T = i), i = 0, . . . , m.
Because the random vectors (Xi , Wi ) satisfy the conditions of Corollary
2 and C(L) is a nonincreasing function of (W1 , . . . , Wn ) it follows from that
Corollary that E[C(L)|T = i] is a nonincreasing function of i. Hence, if the
resulting estimates of E[C(L)|T = i] are not monotone, then an isotonic
regression can be employed to modify them.
A set of components is said to be a cut set if the system is necessarily

failed when all components in this set are failed. A cut set is said to be a
minimal cut set if none of its proper subsets are cut sets. A second simu-
lation approach, which can be used when the number of minimal cut sets
is not too large, uses an identity for Bernoulli sums. Suppose there are r
minimal cut sets, C1 , . . . , Cr , and for the set Ci define an indicator variable
Zi equal to 1 if all the components in Ci fail before time t∗ and equal to 0
Pr
otherwise. Let Z = i=1 Zi , and let
r Y
X
λ = E[Z] = (1 − pj ).
i=1 j∈Ci
Now, it can be shown (see Section 11.3 of 7 for a proof) that for any random
variable R
E[ZR] = λE[R|ZI = 1] (3)

where I is independent of Z, R and is equally likely to be any of the values

1, . . . , r. Now, let
(
0 if Z = 0
R= C(L) .
Z if Z > 0
Because Z = 0 if and only if C(L) = 0, the identity (3 ) yields that
C(L)
E[C(L)] = λ E[ |ZI = 1]. (4)
Z
Using that
P (Zj = 1) aj
P (I = j|ZI = 1) = Pr = Pr
i=1 P (Zi = 1) i=1 ai
where
Y
ai = (1 − pj )
j∈Ci
we can use the preceding to obtain a simulation estimate of E[C(L)] by

performing each simulation run as follows:
a
(1) Generate I such that P (I = j) = Pr j ai , j = 1, . . . , r. Suppose
i=1
I = j.
(2) For i ∈ Cj , generate Wi according to the distribution
Fi∗ (t) = P (Wi ≤ t|Wi ≤ t∗ ) = Fi (t)/Fi (t∗ ) , 0 ≤ t ≤ t∗ .
(3) For i ∈
/ Cj , generate Wi according to the distribution Fi .
(4) Determine Zi , i = 1, . . . , r.
Pr
(5) Determine Z = i=1 Zi .
(6) Determine L.
(7) Determine C(L).
(8) Return the estimator λ C(L) Z .
In contrast to our first estimator, the preceding estimator need not have a
smaller variance than the raw simulation estimator. However, it should be
very efficient when λ is small.
6. Modeling Basket Default Costs

A model for basket default swaps in which a portfolio consists of n assets,
the ith of which defaults at a random time having distribution Fi , was
considered in4 and.2 It was supposed in these papers that if at least r
assets default by a fixed time t∗ then a cost depending on L, the time of

the rth default, was incurred. Thus, the model of 4 is a special case of the
model of the preceding section, in which the system structure is an n− r + 1
of n system which works if and only if at least n− r + 1 of the n components
fail. Letting T be the number of assets that do not fail (i.e., do not default)
by time t∗ , it was suggested in,2 as an improvement on the method of,4 that
E[C(L)] be estimated by continually simulating the system conditional on
the event that T ≤ n − r, and then use the average of the values obtained
for C(L) multiplied by P (T ≤ n − r). That is,2 notes that
E[C(L)] = E[C(L)|T ≤ n − r]P (T ≤ n − r)
and then uses simulation to estimate E[C(L)|T ≤ n − r]. However, because
our approach estimates E[C(L)|T ≤ n − r] by estimating all the quantities
E[C(L)|T = j], j ≤ n − r, while using the known probabilities P (T = j), it
is a stratified version of the estimator of 2 and thus necessarily has a smaller
variance. (Intuitively, there will be a lot more variance in the conditional
distribution of C(L) given that T ≤ n − r, than there will be in the condi-
tional distribution of C(L) given that T = j, for any j ≤ n − r.)
In cases where P (T ≥ r) is small, and nr is not too large, the second

simulation method of the previous section should also be considered.
Example 2. The paper2 gave an example in which there are 10 indepen-

dent exponential random variables Wi , i = 1, . . . , 10, with respective rates
0.03, 0.01, 0.02, 0.01, 0.005, 0.001,
0.002, 0.002, 0.017, 0.003,
and with respective additive costs 0.3, 0.1, 0.2, 0.1, 0.3, 0.1, 0.2, 0.2, 0.1, 0.3
incurred if Wi < t, i = 1, . . . , 10. The method of 2 performed slightly bet-
ter than raw simulation. The following indicates the variance of the raw
estimator and of our estimator for different values of t.
Table 1.
t raw estimator variance variance of our estimator

5 0.021 0.003
10 0.034 0.006
15 0.045 0.008
20 0.052 0.009
25 0.057 0.010
30 0.062 0.011
As can be seen our estimator is far superior to the raw simulation esti-
mator, and thus to the estimator of.2
7. A Round Robin Tournament

In a round robin tournament of n + 1 players, each of the n+1

2 pairs play
a match. The players who win the greatest number of matches are the
winners of the tournament. Suppose that the results from all matches are
independent, and that P (i, j) = 1 − P (j, i) is the probability that i beats
j in their match. Let I be the indicator of the event that player n + 1
is the sole tournament winner, and suppose further that we want to use
simulation to estimate E[I] = P (I = 1). Letting Xi be the indicator of
the event that player n + 1 beats player i in their match, i = 1, . . . , n, an
efficient way to estimate E[I] would be to let T = ni=1 Xi and then do
P
the simulation stratified on T . That is, compute P (T = j), j = 0, 1, . . . , n,

and use
n−1
X
E[I] = E[I|T = j]P (T = j) + P (T = n).
j=[n/2]+1
To estimate E[I|T = j], we would generate X1 , . . . , Xn conditional on T = j

and then generate the outcomes of the n2 games that do not involve player
n + 1, and then take I as the estimator of E[I|T = j] from that run.
Antithetic variables should be effective, so when doing the next run we
should use (in the same manner) the same uniforms (subtracted from 1) that
were just used. As always it is advised to do a small simulation preparatory
study to estimate the conditional variances, so as to set the number of runs
done conditional in each strata in the final study.
Let Wi be the event that player i is the sole tournament winner, and
suppose now that, rather than just wanting to estimate P (Wn+1 ), we want
to estimate P (Wi ) for all i = 1, . . . , n + 1. In this situation, we suggest
a post-stratification technique. Start by solving the n + 1 sets of linear
equations so as to determine the quantities
P (Ti = j), j = 1, . . . , n, i = 1, . . . , n + 1
where Ti is the number of matches that player i wins. Now perform a fixed
number of raw simulations of the tournament. Based on the results, let
N (i, j) be the number of simulation runs in which player i wins exactly j
matches, and let W (i, j) denote the number of simulation runs in which
player i both wins exactly j matches and, in addition, is the sole winner of
the tournament. Now, take W (i, j)/N (i, j) as an estimate of P (Wi |Ti = j).
Since
n
X
P (Wi ) = P (Wi |Ti = j)P (Ti = j)
j=1
this yields our estimate of P (Wi ); namely,

n
X W (i, j)
P (Ti = j).
j=1
N (i, j)
Acknowledgments
The research of Mark Brown was supported by the National Security
Agency, under Grant H98230-06-01-0149.
References
1. Barlow, R. E., Bartholomew, D. J., Bremner, J. M., and Brunk, H. D., (1972),
Statistical Inference under Order Restrictions: Isotonic Regression, Wiley,
New York.
2. Chen, Z. and Glasserman, P., (2008), Fast Pricing of Basket Default Swaps,
Operations Research, 56, 2, 286-303.
3. Efron, B., (1965) Increasing Properties of Polya Frequency Functions. Annals
of Mathematical Statistics, 36, 272-279.
4. Joshi, M. and Kainth, D., (2004), Rapid and Accurate development of Prices
and Greeks for nth to Default Credit Swaps in the Li Model, Quantitative
Finance, Vol. 4, Institute of Physics Publishing, London, UK, 266-275.
5. Keilson, J., and Gerber, H., (1971), Some Results for Discrete Unimodality,
Jour. Amer. Statist. Assc. 66, 386-389.
6. Ross, S. M., (2007), Introduction to Probability Models, Nineth ed., Aca-
demic Press, Burlington, MA.
7. Ross, S. M., (2006), Simulation, fourth ed., Academic Press, Burlington, MA.
May 11, 2011 15:18 WSPC - Proceedings Trim Size: 9in x 6in 06-zeng
115
Filtering with Counting Process Observations and Other

Factors: Applications to Bond Price Tick Data
Xing Hu
Department of Economics
Princeton University
Princeton, 08544, USA
Email: xinghu@princeton.edu
David R. Kuipers
Department of Finance
Henry W. Bloch School of Business and Public Administration
University of Missouri at Kansas City
Kansas City, MO 64110, USA
Email: kuipersd@umkc.edu
Yong Zeng
Department of Mathematics and Statistics
University of Missouri at Kansas City
Kansas City, MO 64110, USA
Email: zengy@umkc.edu
In this paper, we propose an extended filtering micromovement model. The

model captures the two main stylized facts of the bond price tick data: random
trading times and trading noises. In the intrinsic value process for the transac-
tion price of 5-year U.S. Treasury note, we extend the volatility part by adding
the buyer-seller initiation dummy. For the extended model, we present the nor-
malized and un-normalized filtering equations, a robustness theorem and the
consistency of Bayes estimates. Based on the robustness theorem, we employ
the Markov chain approximation method to construct a robust recursive algo-
rithm for computing the posteriors and Bayes estimates. We present a Monte
Carlo example to demonstrate that the computed Bayes estimates converge to
their true values. The algorithm is applied to one and an half month of intraday
transaction prices of 5-year Treasury notes. Bayes estimates are obtained. Es-
pecially, the sign of the buyer-seller initiation dummy is significantly negative,
supporting that the inventory theory dominates in the bond trading.
Keywords: Ultra high frequency data; filtering; counting process; Markov chain
approximation method; Bayes parameter estimation; price discreteness; and
price clustering.
116 X. Hu, D. R. Kuipers and Y. Zeng
1. Introduction
Asset pricing models in financial markets can broadly be classified into two
types based on the observed data frequency: macromovement and micro-
movement models. Macromovement models generally refer to data observed
at daily or less frequent horizons. For example, Figure 1 depicts a time series
plot of daily closing prices for trading in a particular 5-year U.S. Treasury
note in the bond market during the six-week period November 15, 2000
through December 29, 2000. The macromovement for this security consists
of 31 data points for the 31 business days during this time period. Despite
the brevity, the stochastic nature of the time series is evident in the fig-
ure, and is often modelled in the econometrics and mathematical finance
literature by continuous-time diffusion models such as geometric Brownian
motion (GBM), stochastic volatility (SV), jump diffusion, or more elaborate
Markov-class stochastic price processes.
Alternatively, transactional price data on an intraday basis is the do-
main for micromovement asset pricing models. Engle5 refers to data of this
type as ultra high frequency (UHF), because it contains the trading times
and prices for all market activity at the maximum level of disaggregation.
Because traders choose to transact at random times during the trading
day, due to both information-based and liquidity-based motives (O’Hara21 ),
econometricians model the duration between trades as a stochastic phenom-
ena resulting in irregularly-spaced time series. Engle5 develops a general
framework for modelling such time series with many recent developments.
Figure 2 depicts the micromovement data for the same 5-year U.S. Trea-
sury note graphed in Figure 1, during the same time period. There are over
2,200 transactions shown in the figure. The general movement in Figure 2
can be seen as its subsample macromovement from Figure 1, with an overlay
of an additional noise process that results from the UHF data. The source
of this noise process can fundamentally be related to the asynchronicity of
trading in security markets; actual price processes are not continuous-time
realizations, but are observed with discrete sampling, at the irregular inter-
vals where trades take place. Further, markets set minimum price variation
conventions for trading–the so-called “tick size” in the market–which re-
sults in price discreteness noise. In the 5-year Treasury note market, the
minimum tick size is 1/128th percent of par value, so that realized price
processes exhibit some level of noise as prices move from discrete tick to
tick, or level by level.
Finally, even after consideration of asynchronous trading and price dis-
creteness, additional noise is observed in actual security market price pro-
Applications to Bond Price Tick Data 117
Fig. 1.
Daily closing prices of 5−year Treasury Notes from Nov15 2000 to Dec29 2000
104
103.5
103
102.5
price(% of par value)
102
101.5
101
100.5
100
5 10 15 20 25 30
day (31 business days)
cesses due to the tendency for trades and price changes to cluster on certain
ticks and in particular increments compared to others. This price clustering
noise, extensively modelled in the finance literature by Harris15 and others,
can be seen for the 5-year U.S. Treasury note in Figure 3.
Given this background, we reach the simple intuition, prevalent in the
finance and econometrics modelling literature, that micromovement mod-
els should be built upon fundamental intrinsic value price processes, with
an overlay of market microstructure noise derived from any of the several
sources noted above. Zeng25 proposes one such model for asset prices in
a trading market, a filtering micromovement (hereafter, FM) model, that
incorporates UHF data. Corresponding to the macromovement, an unob-
servable intrinsic value process for an asset is assumed and it is the per-
manent component and has a long-term impact on price. Trading times
are assumed to be driven by a conditional Poisson process. Prices are cor-
rupted observations of the intrinsic value process at the trading times by
market microstructure noise, which is explicitly and flexibly modeled by a
random transformation. Comparing with intrinsic value, noise is the tran-
sient component and only has a short-term impact (when a trade happens)
on price. The FM model can be framed as a filtering problem with counting
Fig. 2.
Transaction data of 5−year Treasury Notes from Nov15 2000 to Dec29 2000
104
103.5
103
102.5
price(% of par value)
102
101.5
101
100.5
100
5 10 15 20 25 30
day (31 business days)
process observations. This connection introduces the powerful stochastic fil-

tering theory, which has found great success in engineering and networking.
Then, the unnormalized Duncan-Mortensen-Zakai-like filtering equation
and the normalized, Kushner-Stratonovich (KS) (or Fujisaki-Kallianpur-
Kunita)-like filtering equations are derived. They uniquely characterize the
evolutions of the continuous-time likelihoods and posteriors, respectively.
Moreover, the Bayes estimation via filtering for the intrinsic value process
and the related parameters are developed by employing the Markov chain
approximation method to numerically solve the filtering equation. Further-
more, the Bayesian hypothesis testing or model selection via filtering for
the FM model is developed in Kouritzin and Zeng.17
In the paper, we study and perform Bayes estimation via filtering for an
extended FM model by allowing an observable economic variable to affect
the volatility of the intrinsic value process. The variable is the buyer-seller
initiation dummy, which is observable in the bond market. This is a special
feature of bond market. Such variable is not observable in stock market, but
it can be inferred by some tests and a widely-used test is the Lee and Ready
test (see18 ). By testing whether this dummy is statistically significant, we
can infer whether the information theory or the inventory theory better
Fig. 3.
5−Year Treasury Note,Nov15−Dec20 2002 5−Year Treasury Note,Nov15−Dec20 2002
800 1200
700
1000
600
800
500
frequency
frequency
400 600
300
400
200
200
100
0 0
−40 −20 0 20 40 1/32 1/128 1/64
price changes (1/128th) fractional parts
fit the bond market. Moreover, the trading times are assumed to follow
an Exponential Autoregressive Duration (EACD) model proposed by Engle
and Russel6 instead an inhomogeneous Poisson process. The EACD model
is estimated and tested. We find that the EACD model fits the trading
times pretty well.
To the best of our knowledge, Frey11 and Frey and Runggaldier12 are
the first papers that employ the non-linear filtering technique to model
UHF data. Their viewpoint is to model the unobserved volatility process,
which is crucial for option pricing. Their model is able to capture the Pois-
son random arrival times in UHF data. Cvitnic, Lipster and Rozovskii4
extends the previous model to a more general framework. However, market
microstructure noise is missing in these models and they did not consider
the impact of other observable variable on volatility.
In Section 2, we present the extended FM model with an observable
variable in two equivalent fashions. In Section 3, we present the continuous-
time Bayes estimation via filtering for the extended FM model including
the filtering equations, a robustness theorem, a recursive algorithm, and a
theorem on the consistency of the Bayes estimates. In Section 4, we present
a Monte Carlo example. In Section 5, we provide empirical results for the 5
year Treasury note tick data previously shown. We conclude in Section 6.
2. The Simple FM Model with an Observable Factor

We present the extended FM model in two equivalent representations and
show they are equivalent.
2.1. Representation I: Constructing Price from Intrinsic

Value
Based on the simple intuition obtained in Section 1 that the bond price is
formed from an intrinsic value by incorporating the market microstructure
noises that arise from the trading activities, we construct an FM model
with an observable economic factor - the buyer-seller initiation dummy.
In general, there are three steps in constructing the tick-by-tick bond
price process Y from the intrinsic value process X. Since the intrinsic value
can only be partially observed through prices, (X, Y ) forms a partially-
observed system. In order to prepare for parameter estimation, we would
like to add in another process θ. To further allow other observable economic
factor, we would like to include an observable vector process V . Therefore,
we would like to consider an enlarged partially-observed model (θ, X, Y, V ).
Assume (θ, X, Y, V ) is defined in a complete probability space (Ω, F , P )
with a filtration {Ft }0≤t≤∞ satisfying the usual conditions (see Protter22 ).
Step 1: We consider a specific model for the 5-year bond’s intrinsic
value process, which is assumed as below:
dXt
= µdt + (σ + κVt )dBt (1)
Xt
where Bt is a standard Brownian motion and Vt is the observable buyer-
seller initiation dummy defined as below. We assume trading times are
denoted by t1 , t2 , . . . , ti , . . ..
V (t) = V (ti ) if ti ≤ t < ti+1 ,
where V (ti ) takes value 1 if a trade is buyer-initiated (namely, the trade

is a“Take”, take the ask quote to sell); and takes value 0 if seller-initiated
(namely, the trade is a “Hit”, hit the bid quote to buy); and the value of
the dummy stays the same until next trade. The instantaneous expected
return is µ. The instantaneous volatility is σ + κVt . When κ is positive
(negative), the volatility increases (decreases) from σ to σ + κ for a buyer-
initiated trade. Modelling the volatility structure in this fashion captures
the possibility that price moves in the market exhibit asymmetric volatil-
ity as a function of transacting dealers’ inventory position (Fleming and
Rosenberg,10 ).
In the general case considered in Hu, Kuipers and Zeng,16 a mild as-
sumption on (θ, X), stated below, is invoked so that all relevant stochastic
processes are included.
Assumption 1. (θ, X) is the unique solution of a martingale problem for

a generator Av such that for a function f in the domain of Av ,
Z t
Mf (t) = f (θ(t), X(t)) − Av f (θ(s), X(s))ds
0
is a Ftθ,X,V -martingale, where Ftθ,X,V is the σ-algebra generated by

(θ(s), X(s), V (s))0≤s≤t .
The generator and martingale problem approach originated by Strook

and Varadhan in a series of papers and their book23 (see also Ethier and
Kurtz7 for an excellent book on this topic) furnishes a powerful tool for the
characterization of Markov processes. Assumption 1 includes all relevant
stochastic processes such as diffusion, jump, regime-switching, spike, and
their combinations for modeling asset price.
Assumption 1 is more general than that in Zeng25 by allowing other ob-
servable economic factors to affect (θ, X), namely, allowing other observable
factors in the generator, Av .
Let θ = (µ, σ, κ, ρ) where ρ is a parameter for non-clustering noise to
be described. The generator of the 5-year bond’s intrinsic value process
described in Equation (1) is
∂ 1 ∂2
Av f (θ, x) = µx f (θ, x) + (σ + κv)2 x2 2 f (θ, x). (2)
∂x 2 ∂x
However, in UHF data, the price can not be observed continuously in
time, neither can it moves continuously. Therefore, we need two more steps.
Step 2 takes care of the trading times and Step 3 takes care of the trading
noise.
In Step 2: We assume trading times t1 , t2 , . . . , ti , . . ., are driven by an
Exponential Autoregressive Conditional Duration (EACD) model proposed
by Engle and Russell.6
We use a couple ways to describe EACD model. The first way is from
the viewpoint of point process. We view trading times t1 , t2 , . . . , ti , . . . as a
conditional Poisson process specified by its stochastic intensity

X
−1
λ̄(t) = I{ti < t ≤ ti+1 }(t)ψi+1 (3)
i≥0
where ψi+1 is the conditional expectation of the i + 1 duration given the

past information and ψi+1 is given in Equation (7).
In the other way, EACD model is specified in terms of the conditional
density of the durations. Let ∆ti+1 = ti+1 − ti be the time interval between
two trades, which is called duration. We will specify the density of ∆ti+1
directly. Let η be a vector of parameters in EACD model. Recall
E(∆ti+1 |∆i , ..., ∆1 ; η) = ψ(∆i , ..., ∆1 ; η) = ψi+1 . (4)
Then, the EACD class of models consists of parameterizations of ψi+1 and

the assumption that
∆i+1 = ψi+1 εi+1 (5)
where
{εi } ∼ i.i.d. with exponential distribution of mean one. (6)
A general specification of ψi+1 , which is similar to ARMA model in time

series literature to accommodate the unlimited past memory characteristic
of ∆t, is
m
X q
X
ψi+1 = ω + αj ∆i+1−j + βk ψi+1−k . (7)
j=1 k=1
Equations (4) - (7) consist of an EACD(m, q) model where the m and q

refer to the orders of the lags.
In Step 3: Y (ti ), the price at time ti , is corrupted from X(ti ), the
intrinsic value, with trading noise. We model noise by a random transfor-
mation, y = F (x), with the transition probability p(y|x). Namely,
Y (ti ) = F (X(ti )).
The random transformation, F (x), is flexible and is able to accommodate

different kinds of noise. As demonstrated in Section 1, there are three im-
portant types of noise that have been identified in bond tick price : dis-
crete, clustering, and nonclustering. First, intraday prices move discretely,
resulting in “discrete noise”. Second, because prices do not happen evenly
on all 128ths ticks, but more concentrate on some coarser ticks such as
odd 64ths and 32ths or coarser, “price clustering” is obtained. Third, the
“non-clustering noise” contains all other unspecified noise. Below we fol-

low the approach in Zeng25 to construct F (x) to accommodate these three
noise. There are three steps. For simple notation, at a trading time ti , let
1
x = X(ti ), y = Y (ti ), and y 0 = Y 0 (ti ) = R[X(ti ) + Wi , M ], where Wi is to
be defined as the non-clustering noise.
• Step (a): Add non-clustering noise W ; x0 = x+W , where W is the

non-clustering noise at trade i. We assume {Wi }, are independent
of the value process, and they are i.i.d. with a doubly geometric
distribution:

(1 − ρ) if w = 0
P {W = w} = 1 128|w| 1 2 .
2 (1 − ρ)ρ if w = ± 128 , ± 128 ,···
• Step (b): Incorporate discrete noise by rounding off x0 to its closest
1
tick, y 0 = R[x0 , 128 ].
• Step (c):Incorporate clustering noise by biasing y 0 through a ran-
dom biasing function bi (·) at trade i. {bi (·)} is assumed independent
of {yi0 }. To be consistent with the 5-year bond price data analyzed,
we construct a simple random biasing function only for the tick
of 1/128 percentage. For other tick size, it can be done similarly.
The data to be fitted has this clustering phenomenon: odd thirty-
seconds or coarser are most likely and have about the same frequen-
cies; odd sixty-fourths are the second most likely and have about
the same frequencies; and odd one hundred and twenty-eighths are
least likely and have about the same frequencies. To generate such
clustering, a random biasing function is constructed based on the
following rules: if the fractional part of y 0 is an even one hun-
dred twenty-eighths, then y stays on y 0 with probability one; if the
fractional part of y 0 is an odd one hundred twenty-eighths, then
y stays on y 0 with probability 1 − α − β, y moves to the closest
odd sixty-fourth with probability α, and moves to the closest odd
thirty-seconds or coarser with probability β.
In brief,
1
Y (ti ) = bi (R[X(ti ) + Vi , ]) = F (X(ti )).
M
The detail of bi (·), and the explicit p(y|x) for F can be found in Appendix
A. A simulation example in Section 4 demonstrates that the constructed
F (x) are able to capture the tick-level sample characteristics of bond price
data.
Before moving to the second representation of the FM model, we give a

couple remarks for this representation about its connections to other models
in finance and statistics literature.
Remark 1. Under this construction, information influences X(t), the value

of an asset, and has a permanent impact on the price while noise modeled by
F (x) (or p(y|x)) only has a transitory impact on price. The formulation is
analogous to the time series VAR structural models used in many market
microstructure papers (see a survey paper13 by Hasbrouck and a recent
paper Hasbrouck14 ).
Remark 2. Furthermore, the formulation is closely connected to the two-

scale frameworks combining market microstructure noises in recent liter-
ature of realized volatility estimators. See Zhang, Mykland and Ait- Sa-
halia,27 Ait-Sahalia, Mykland and Zhang,1 Bandi and Russell,2 and Fan
and Wang.8 Especially, Li and Mykland in20 demonstrates that rounding
noise in UHF data may seriously distort even the two-scale estimators of re-
alized volatility, and the error could be infinite. Note that price discreteness
exists in the bond price data also.
2.2. Filtering with Counting Process Observations

Because of bond price discreteness, we can formulate the bond prices as a
collection of counting processes in the following form:
Y1 (t)
 
 Y2 (t) 
~ (t) = 
Y  . ,

(8)
 .. 
Yn (t)
where Yj (t) is the counting process recording the cumulative number of

trades that have occurred at the jth bond price level (denoted by yj ) up
to time t. We make additional four more mild assumptions on the model
so that both representations are equivalent. We assume that there exists a
reference measure Q with certain properties.
Assumption 2. Under the reference measure Q, {Yj }nj=1 are independent

unit Poisson processes.
Assumption 3. Under the reference measure Q, (θ, X), Y1 , Y2 , . . . , Yn are

independent.
Assumption 4. The intensities are of the form: λj (t) = λ̄(t)p(yj |x), where
P −1
λ̄(t) = i I{ti < t ≤ ti+1 }(t)ψi+1 is the total trading intensity at time t
and p(yj |x) is the transition probability from x to yj , the jth price level.
This assumption imposes a desirable structure for the intensities of the

model. It means that the total trading intensity λ̄(t) determines when the
next trade will occur and p(yj |x), which is the same as p(y|x) of y = F (x),
determines at which price level the next trade will occur given the value is
x. Note that p(yj |x) models how the trading noise enters the price process.
The intensity structure plays an essential role in the equivalence of the two
approaches of modeling.
Under probability measure P , Yj (t) is a conditional Poisson process
θ,X,Y
with the stochastic intensity, λj (t) = λ̄(t)p(y
R t j |x). Given
R t Ft , Yj (t) has
a Poisson distribution with parameter 0 λj (s)ds = 0 λ̄(s)p(yj |X(s))ds.
Rt Rt
Moreover, Yj (t)− 0 λj (s)ds is a Ftθ,X,Y - martingale and 0 λj (s)ds is called
the compensator of Yj (t) for each j.
RT
Assumption 5. 0
E P [λ̄(s)]ds < ∞, for t > 0.
Under very mild conditions, the reference measure Q exists. Assumption

5 is a mild technical condition to ensure the existence of Q. The Girsanov
type change of measure theorem for Poisson process can be found in (T2
and T3 Theorems in Chapter 4.2 of Bremaud3 ).
Remark 3. Under this representation, (θ, X) becomes the signal, which

cannot be observed directly, but can be partially observed through the
counting processes, Y~ with the observable factor, V . The counting pro-
~ (t) is distorted observations of X(t) by trading noise,
cess observations Y
modeled by p(yj |x). Hence, (θ, X, Y~ , V ) is framed as a filtering model with
counting process observations and an observable factor.
Remark 4. To solve problems in mathematical finance such as the option

pricing and hedging and the portfolio selection, the stochastic differential or
integral equation form of the most recent price is needed. However, the pre-
vious two representations do not provide such a form of price. And a third
representation is given in Lee and Zeng19 and Xiong and Zeng.24 In those
two papers, the option pricing and hedging through local risk minimizing
criterion and the portfolio selection problem are studied respectively.
2.3. The Equivalence of the Two Representation

We construct the price from the intrinsic value in Representation I and
frame it as a filtering problem with counting process observations in Rep-
resentation II. The following proposition states their equivalence in dis-
tribution. This guarantees the statistical inference based on the second
representation is also valid to the first one.
Proposition 2.1. The two representations of the model in Sections 2.1

and 2.2, respectively, have the same probability law.
The proof is exactly the same as that of the proposition in,26 where
the main idea is to show both representations have the same stochastic
intensity kernel.
Since the observed prices can be equivalently represented by Y or Y ~
~
in Representations I or II, respectively, we use Y or Y in exchange of one
another in the rest of the paper.
3. Bayes Estimation via Filtering

This section presents the filtering equations for the FM model with an ob-
servable factor, a robustness theorem, and a theorem for the consistency of
the Bayes estimates without proofs. A more general version of these results
with proofs can be found in Hu, Kuipers and Zeng.16 For the FM model
considered in this paper, we construct an efficient recursive algorithm to
compute the joint posteriors and the Bayes estimates. The robustness theo-
rem not only provides a blueprint through the Markov chain approximation
method to construct recursive algorithms, but also ensures the robustness
of such algorithms.
3.1. The Statistical Foundations

We first study the continuous-time joint likelihood, the likelihood function
(from frequentists’ viewpoint), the integrated likelihood (from Bayesians’
viewpoint), and the posterior of the proposed model. They are character-
ized by the unnormalized and normalized filtering equations. For those who
are interested in the continuous-time likelihood ratio and Bayes factors for
hypotheses testing or model selections, and the system of evolution equa-
tions characterizing the evolution of Bayes factors, we refer to Kouritzin
and Zeng17 for details.
3.1.1. The Continuous-time Joint Likelihood

The probability measure P of (θ, X, Y ~ , V ) can be written as P = Pθ,x,v ×
Py|θ,x,v , where Pθ,x,v is the probability measure for (θ, X, V ) such that
Mf (t) in Assumption 1 is a Ftθ,X,V -martingale, and Py|θ,x,v is the con-
ditional probability measure on DRn [0, ∞) for Y ~ given (θ, X, V ) (where
DRn [0, ∞) is the space of right continuous with left limit functions). Un-
der P , Y~ relies on (θ, X, V ). Recall that there exists a reference measure
Q such that under Q, (θ, X, V ), Y ~ become independent, (θ, X, V ) remains
the same probability law and Y1 , Y2 , . . . , Yn become unit Poisson processes.
Therefore, Q can be decomposed as Q = Pθ,x,v × Qy , where Qy is the prob-
ability measure for n independent unit Poisson processes. One can obtain
the Radon-Nikodym derivative of the model, that is the joint likelihood of
~ , V ), L(t), as (see3 pg 166),
(θ, X, Y
dP dPθ,x,v dPy|θ,x dPy|θ,x,v
L(t) = (t) = (t) × (t) = (t)
dQ dPθ,x,v dQy dQy
n
Y
(Z
t Z th i
) (9)
= exp log λj (s−)dYj (s) − λj (s) − 1 ds ,
j=1 0 0
or in stochastic differential equation(SDE) form:

Xn Z th i
L(t) = 1 + λj (s−) − 1 L(s−)d(Yj (s) − s).
j=1 0
~
3.1.2. The Continuous-time Likelihoods of Y
Note that X are not observable. To do parameter estimation, we would like
to have the likelihood of Y ~ alone. Namely, we would like to integrate out
X. The probability way to achieve this is to use conditional expectation.
Let FtY,V = σ{(Y ~ (s), V (s))|0 ≤ s ≤ t} be all the available information up
to time t. We use E Q [X] and E P [X] to indicate that the expectation is
taken with respect to the measures Q and P , respectively.
Definition 1. Let
Z
ρ(f, t) = E Q [f (θ(t), X(t))L(t)|FtY,V ] = f (θ, x)ρt (dθ, dx).
In the frequentists’ paradigm, (θ(0), X(0)) is fixed and the likelihood of

Y is E Q [L(t) | FtY,V ] = ρ(1, t). In Bayesian paradigm, we assume a prior on
(θ(0), X(0)) and the integrated (or marginal) likelihood of Y is also ρ(1, t).
3.1.3. The Continuous-time Posterior

In order to do Bayes estimation, we would like to consider the conditional
distribution and the posterior.
Definition 2. Let πt be the conditional distribution of (θ(t), X(t)) given
FtY,V and let
Z
P Y,V
π(f, t) = E [f (θ(t), X(t))|Ft ] = f (θ, x)πt (dθ, dx).
In Bayesian paradigm, we assume a prior on (θ(0), X(0)), so πt becomes

the continuous-time posterior, which is determined by π(f, t) for all continu-
ous and bounded f . The relationship between ρ(f, t) and π(f, t) is provided
in Bayes Theorem (see,3 page 171), which states: π(f, t) = ρ(f, t)/ρ(1, t).
For this reason, the equation governing the evolution of ρ(f, t) is called the
unnormalized filtering equation, and that of π(f, t) is called the normalized
filtering equation.
3.2. Filtering Equations

Stochastic partial differential equations (SPDEs) provide an powerful ma-
chinery to characterize the infinite dimensional continuous-time likelihoods
and posteriors. The likelihoods are characterized by the unnormalized filter-
ing equation and the posteriors are characterized by the normalized filtering
equation. The following theorem put together these two filtering equations.
Theorem 1. Suppose that (θ, X, Y~ , V ) satisfies Assumptions 1 - 5. Then,
ρt is the unique measure-valued solution of the SPDE, the unnormalized
filtering equation,
Z t n Z t
X
ρ(f, t) = ρ(f, 0)+ ρ(Af, s)ds+ ρ((λ̄pj −1)f, s−)d(Yj (s)−s), (10)
0 j=1 0
for t > 0 and f ∈ D(A), the domain of generator A, where λ̄ = λ̄(t) is

the trading intensity and is assumed to be Ftθ,X,Y,V -predictable, and pj =
p(yj |x) is the transition probability from x to yj .
πt is the unique measure-valued solution of the SPDE, the normalized
filtering equation,
Z t
π(f, t) = π(f, 0) + π(Af, s)ds
0
n Z th (11)
X π(f λ̄pj , s−) i
+ − π(f, s−) d Yj (s) − π(λ̄pj , s)s .
j=1 0
π( λ̄p j , s−)
Moreover, when the trading intensity is λ̄(t) is only FtY,V -predictable,

namely, λ̄(t) depends on only the past observations and the observable
factors, the normalized filtering equation is simplified as
Z t n Z th
X π(f pj , s−) i
π(f, t) = π(f, 0) + π(Af, s)ds + − π(f, s−) dYj (s).
0 j=1 0
π(pj , s−)
(12)
Remark 5. Note that λ̄(t) disappears in Equation (12). This not only
greatly reduces the computation in of Bayes estimates, but also implies
that the joint posterior and the Bayes estimates are independent or ”model
free” of the assumptions of trading times as long as λ̄(t) is FtY,V -predictable.
Remark 6. For the FM model studied in this paper, the trading intensity
is given by Equation (3), which is, indeed, FtY,V -predictable. Therefore,
we only use the simplified filtering equation (12) to construct a recursive
algorithm to compute Bayes estimates.
Another nice property of the filtering equations for πt is that they sat-
isfy a separation principle in the following sense. Let the trading times be
t1 , t2 , . . . , then Equation (12) can be written in two parts. The first de-
scribes the evolution without trades and is called the propagation equation.
The second describes the update when a trade occurs and is called the
updating equation. The propagation equation is written as
Z ti+1 −
π(f, ti+1 −) = π(f, ti ) + π(Af, s)ds. (13)
ti
It has no random component and this implies that the posterior evolves
deterministically between trades.
When the price at time ti+1 happens at the jth price level, the updating
equation is written as
π(f pj , ti+1 −)
π(f, ti+1 ) = . (14)
π(pj , ti+1 −)
Note that the price level j, the observation, is random. Therefore, the up-
dating equation is random.
3.3. A Convergence Theorem and Recursive Algorithms

Theorem 1 describes the evolutions of the continuous-time likelihoods and
posteriors. They are both infinite dimensional. This means, in order to com-
pute them, one needs to approximate the infinite dimensional system by a
finite dimensional system, based on which algorithms are constructed. The

algorithms, based on the filtering equations, are naturally recursive, han-
dling one datum at a time. Moreover, the algorithms are easily parallelizable
and thus can make real-time updates and handle large data sets. One ba-
sic requirement for the recursive algorithms is robustness (or consistency):
The approximate likelihoods or posteriors, computed by the recursive al-
gorithms, converges to the true ones. The following theorem ensures the
robustness and provides a blueprint for constructing consistent algorithms
through Kushner’s Markov chain approximation methods.
Let (θ , X ) be an approximation of (θ, X). Then, we define
Y,1 (t)
 
 Y,2 (t) 
Y~ (t) =  .  , (15)
 
 . .
Y,n (t)
Y ~
where Y,j (t) has the stochastic intensity
λ̄(t)p(yj |X (t), θ ).
Let Ft =
~ (s), 0 ≤ s ≤ t), and L (t) = L θ (s), X (s), Y (s)

σ(Y 0≤s≤t
as in Equa-
tion (9). We use the notation, X ⇒ X, to mean X converges weakly to
X in the Skorohod topology as → 0. We assume that (θ , X , Y ~ , V ) are
on (Ω , F , P ), and Assumptions 1 - 5 are also satisfied for (θ , X , Y ~ , V ).
Then, a reference measure Q exists with similar properties. Next, we de-
fine the approximations of ρ(f, t), and π(f, t) and state the theorem of
robustness.
Definition 3. Let
h ~
i
ρ (f, t) = E Q f θ (t), Xx (t) L (t)|FtY ,

and
h ~ i
π (f, t) = E P f θ (t), Xx (t) |FtY .
Theorem 2. Suppose that Assumptions 1 - 5 hold for the models

(θ, X, Y~ , V ) and that Assumptions 1 - 5 hold for the approximate models
~ , V ). Suppose (θ , X ) ⇒ (θ, X) as → 0.
(θ , X , Y
Then, as → 0, (i) Y ~ ⇒ Y~ under P and P ;
for all bounded continuous functions, f
(ii) ρ,c (f, t) ⇒ ρ(f, t); and (iii) π,c (f, t) ⇒ π(f, t).
Remark 7. Part (i) implies the convergence of the approximate observa-

tions to the true ones under the physical measures; Part (ii) implies the
robustness of the (integrated) likelihood; Part (iii) implies the robustness

of posterior.
This theorem offers a three-step blueprint for constructing a robust re-

cursive algorithm based on Kushner’s Markov chain approximation method
to compute the continuous-time likelihoods or posteriors. Since we focus
on Bayes estimation in this paper, here, we first give a big picture as
to how to construct such an algorithm for computing the posterior and
Bayes estimates for a model. There are three steps. In Step 1, we produce
(θ , X ), the Markov chain approximation to (θ, X). Note that (θ , X ) is
restricted to the discrete finite-dimensional state space of (θ , X ). We also
obtain p,j = p(yj |θ , x ) as an approximation to pj = p(yj |θ, x). In Step
2, we obtain the filtering equation for π (f, t) for the approximate model
(θ , X , Y , V ) by applying Theorem 1. Since the FM model considered in
this paper has a trading intensity depending only on the past of observa-
tions, we only consider the simplified normalized filtering equation. Recall
the separation principle and we can write the filtering equation for the
approximate model as the propagation equation:
Z ti+1 −
π (f, ti+1 −) = π (f, ti ) + π (A f, s)ds, (16)
ti
and the updating equation (assuming that a trade at jth price level occurs
at time ti+1 ):
π (f p,j , ti+1 −)
π (f, ti+1 ) = . (17)
π (p,j , ti+1 −)
In Step 3, we turn Equations (16) and (17) to the recursive algorithm in
discrete state space and in discrete times by two substeps: (a) expresses
π (·, t) as a finite array using the components π (f, t) with lattice-point
indicator f and (b) use Euler scheme to approximate the time integral in
(16). Details of the construction of the algorithm are given in the sub-section
below.
3.4. The Recursive Algorithms for Posteriors and Bayes

Estimates
In the FM model considered in this paper, we can classify the parameters
into three groups. The parameters of clustering noise, α and β is Group I
and can be estimated through the method of relative frequency. The param-
eters of EACD model for the trading times are Group II and they can be
estimated using the maximum likelihood estimation proposed by Engle and
Russell.6 The other parameters, µ, σ, κ and ρ, are Group III and they are
estimated by Bayes estimation via filtering through the recursive algorithm
to be constructed.
We follow the three-step blueprint for the Markov chain approximation
method summarized in the end of Section 3.3 to construct the recursive al-
gorithms. Finally, we show the consistency of the algorithms. For notational
simplicity, let ~θ~ = (µµ , σσ , κκ , ρρ ) to denote an approximate discretized
parameter signal, which is random in Bayesian framework.
3.4.1. Step 1: Construct (~θ~, Xx )

First, we latticize the parameter spaces of µ, σ, κ, ρ and the state space of
X. Suppose there are nµ + 1, nσ + 1, nκ + 1, nρ + 1 and nx + 1 lattices in
the latticized spaces of µ, σ, κ, ρ and X respectively. For instance,
µ : [αµ , βµ ] → {αµ , αµ + µ , . . . , αµ + (nµ − 1)a , βa }
where the number of lattices is nµ + 1. Define µk = αµ + kµ , the kth
element in the latticized parameter space of µ, and define σl , κm , ρn and
Xw similarly. Let
~θ~v = (µk , σl , κm , ρn )
as where ~v = (k, l, m, n).

We construct a birth and death generator Aε,v , such that Aε,v → Av
pointwisely. Namely, we construct a birth and death process (~θ~, Xx ), a
simple example of Markov chain, to approximate (θ, X(t)) using the gener-
ator for the GBM with observable factor. Recall the generator in Equation
(2), which consists the first and second derivatives. We employ the central
difference approximation to the differentials to construct the approximate
generator as below:
Aε f (~θ~v , xw )
!
f (~θ~v , xw + x ) − f (~θ~v , xw − x )
= µk xw
2x
!
1 f (~θ~v , xw + x ) + f (~θ~v , xw − x ) − 2f (~θ~v , xw )
+ (σl + κm v)2 x2w
2 (x )2
= β(~θ~v , xw )(f (~θ~v , xw + x ) − f (~θ~v , xw ))
+ δ(~θ~v , xw )(f (~θ~v , xw − x ) − f (~θ~v , xw )), (18)

where
(σ + κv)2 x2w

1 µk xw
β(~θ~v , xw ) = 2
+ ,
2 x x
and
(σ + κv)2 x2w

1 µk xw
δ(~θ~v , xw ) = 2
− .
2 x x
β(~θ~v , xw ) and δ(~θ~v , xw ) are the birth and death rates, respectively, and
should be nonnegative. If necessary x can be made small to ensure the
nonnegativity.
Clearly, Aε,v → Av . So, (~θ~, Xx ) ⇒ (~θ, X) as ε → 0 where ε =
max(µ , σ , κ , ρ , x ). With the approximate model (~θ~, Xx (t)), we have
the approximate Y ~ε which is defined by Equation (15).
The counting process observations can be treated as Y ~ (t) defined by
~
Equation (8) or Yε (t) defined by Equation (15) conditioning on whether the
driving process is (~θ, X(t)) or (~θ , Xx (t)). In the true model, the parame-
ters and the bond value are (~ θ, X(t)), and the counting process observations
of bond price are regarded as Y ~ (t). In the approximate model where we in-
tend to compute the posterior and Bayes estimates, we use (~θ~, Xx (t)) to
approach (~θ, X(t)) and the counting process observations of bond price
are regarded as Y ~ε (t). The recursive algorithm is to calculate the joint
posterior and Bayes estimates of the approximate model (~θ~ , Xx , Y ~ε , V ),
which is close to the joint posterior and Bayes estimates of the true model
(~θ, X, Y
~ , V ), by Theorem 2, when ε is small.
3.4.2. Step 2: Obtain the SPDEs of the Approximate Model

When (~θ, , X) is approximated by (~θ~, Xx ), Av by Aε.v , and Y by Yε , there
accordingly exist probability measures Pε and Qε , which approximate P and
Q. It can be checked that Assumptions 1 - 5 hold for (~θ~Xx , Yε ) satisfying
the conditions of Theorem 1. Let (~θ~, Xx ) denote the discretized random
signal and (~θ~v , xw ) a lattice point there.
Definition 4. Let πε,t be the conditional probability mass function of

~
(θ~~, Xx (t)) on the discrete state space given FtYε . Let
~
πε (f, t) = E Pε [f (~θ~, X (t)) | FtYε ] = f (~θ~v , xw )πε,t (~θ~v , xw ),
X
~~
θ v ,xw
where the summation goes over all lattices in the discretized state spaces.
Then, the normalized filtering equation for the approximate model is

given by Equations (16) and (17).
3.4.3. Step 3: Convert to the Recursive Algorithm

We convert Equations (16) and (17) to the recursive algorithm for comput-
ing the approximate joint posterior. First, we define the posterior that the
recursive algorithm calculates.
Definition 5. The posterior of the approximate model at time t is denoted

by
pε (~
θ~v , xw ; t) = πε,t {~θ~ = ~θ~v , X (t) = xw }.
There are two substeps. The key of the first substep is to specify f as
the following lattice-point indicator function:
I{θ~~=θ~~v ,X (t)=xw } (~θ~, X (t)). (19)
We observe:

πε β(θ~~ , X (t))I{θ~ ~ (θ~~ , X (t) + x ), t = β(θ~~v , xw−1 )pε (θ~~v , xw−1 ; t).
=θ~
~ v ,X (t)+x =xw }
Along with suchlike results, Equation (16) turns into

Z ti+1 −
~ ~
pε (θ~v , xw ; ti+1 −) = pε (θ~v , xw ; ti ) + β(~θ~v , xw−1 )pε (~θ~v , xw−1 ; t)
ti

−(β(~θ~v , xw ) + δ(~ θ~v , xw ))pε (~θ~v , xw ; t) + δ(~θ~v , xw+1 )pε (~θ~v , xw+1 ; t) dt.
(20)
Suppose that a trade at jth price level happens at time ti+1 , the updat-
ing Equation (17) becomes,
pε (~θ~v , xw ; ti+1 −)p(yj |xw , ρn )

pε (~
θ~v , xw ; ti+1 ) = P , (21)
0 pε (θ~
~v0 , xw0 ; ti+1 −)p(yj |xw0 , ρn0 )
v 0 ,w
~
where the summation goes over the total discretized space, and p(yj |xw , ρn ),
which is the transition probability from xw to yj , is given by Equation (A.2)
in Appendix A.
In the second substep, we use Euler scheme to approximate the time
integral in Equation (20) in order to obtain a recursive algorithm discrete
in time. Since the probability of the event that two or more jumps occur at
the same time is zero, there are two possible cases for the length of duration.
Case 1, when ti+1 − ti ≤ LL, the length controller in the Euler scheme, we
approach p(~θ~v , xw ; ti+1 −) as
h
p(~θ~v , xw ; ti+1 −) ≈ p(~θ~v , xw ; ti ) + β(~θ~v , xw−1 )p(~θ~v , xw−1 ; ti )
− β(~θ~v , xw ) + δ(~θ~v , xw ) p(~θ~v , xw ; ti )

(22)
i
+ δ(~θ~v , xw+1 )p(~θ~v , xw+1 ; ti ) (ti+1 − ti ).
Case 2, when ti+1 − ti > LL, we select a finer partition {ti,0 =

ti , ti,1 , . . . , ti,n = ti+1 } of [ti , ti+1 ] such that maxj |ti,j+1 − ti,j | < LL and
then approximate p(~θ~v , xl ; ti+1 −) by applying repeatedly Equation (22)
from ti,0 to ti,1 , then ti,2 ,. . ., until ti,n = ti+1 .
The recursive algorithm consists of Equations (21) and (22). Using them,
we calculate the approximate posterior at time ti+1 for (~θ, X(ti+1 )) based
on the posterior at time ti . At time ti+1 , the Bayes estimates of θ~ and
X(ti+1 ) are the expected values of the corresponding marginal posteriors.
To complete the algorithm for posterior, we choose a reasonable prior.
Assume independence between X(0) and θ. ~ The prior for X(0) can be set
by P {X(0) = Y (t1 )} = 1 where Y (t1 ) is the first trade price of a data set
because they are very close. For other parameters, we can simply assume a
uniform prior over the discretized state space of ~θ. At t = 0, we select the
prior as below:
(
1
if xw = Y (t1 )
p(~θ~v , xw ; 0) = (1+nµ )(1+nσ )(1+nκ )(1+nρ ) .
0 otherwise
Finally, we mention two statistical and computational concerns for a

prior on a parameter: suitable range and mesh size. Usually, the marginal
posterior of a parameter obtained from a large data set is concentrated
on a small area around the true value. Then, the uniform prior set on
the small area is sufficient, because the posterior outside is of very small
probability. After having a suitable range, we may choose a suitable mesh
size, which ideally produces a posterior with a unique modal and bell-
shaped distribution as shown in Table 5.1 of.25
3.4.4. Robustness of the Recursive Algorithms

There are two approximations in our recursive algorithms to compute the
posteriors. One is to approximate the integral in the propagation equations
by Euler scheme, whose convergence is well-known. The other one, which is
more important, is the approximation of Equation (12) by Equations (16)
and (17). Since, (~θ~, X ) ⇒ (~θ, X) by construction, Theorem 2 warrants
these convergence in the sense of the weak convergence in the Skorohod

topology, that is, the robustness of the posteriors and Bayes estimates.
3.4.5. Consistency of the Bayes Estimates

Theorem 3. Under the setup of Section 2.1, suppose that the clustering
parameters α, β are known, and (µ, σ, κ, ρ) has a prior. We assume R t that
there are both infinite i such that V (ti ) = V (ti+1 ) = 0 or 1. If µt − 21 0 (σ +
~
κVs )2 ds > 0 for any t > 0, then E[f (µ, σ, κ, ρ)|FtY ] → f (µ, σ, κ, ρ) a.s. as
t → ∞ for any bounded continuous function f .
Rt
The condition “µt − 12 0 (σ + κVs )2 ds > 0” keeps X(t) strictly positive,
and rules out bankruptcy.
Theorem 3 together with the robustness of the recursive algorithm im-
plies that the computed Bayes estimates will converge to their true values.
This is confirmed by the Monte Carlo results presented next.
4. A Monte Carlo Example

We provide a Monte Carlo example for two purposes: One is to show the tick
sample characteristics are captured by the noise modeling and the other is
to show the Bayes estimates of (µ, σ, κ, rho) are consistent. The latter also
serves to test the correctness of the software developed.
4.1. Tick Characteristics of Monte Carlo Data

For simulation, we choose the parameter values close to the estimates from
the 5-year Treasury note data. Let µ = 6.32 × 10−8 , σ = 1.33 × 10−5 ,
κ = −2.71×10−6, (the above three parameters are in per second) ρ = 0.023,
α = 0.250, and β = 0.385. Since λ̄(t) has no impact in estimation and noise,
we assume the trading intensity is constant: a(t) = 0.06 for all t > 0(i.e.,
one trade in about 1/0.06 = 450 seconds). Using these parameters, we
simulated 2,200 observations.
Figure 4 shows the pair of histograms of price changes and of the three
ticks: odd 1/128, odd 1/64 and odd 1/32 or coarser. The similarity of Fig-
ures 4 and 3 shows the proposed modeling of noises is successful to capture
the tick sample characteristics of the 5yr Treasury note price.
4.2. Bayes Estimates for Monte Carlo Data

Having constructed the recursive algorithm in Section 3.3 for calculating the
Bayes estimates, we develop a Fortran program for the recursive algorithm
Fig. 4.
Simulated Data Simulated Data
600 1200
500 1000
400 800
frequency
frequency
300 600
200 400
100 200
0 0
−40 −20 0 20 40 1/32 1/128 1/64
price changes (1/128th) fractional parts
~ X(t)), their
to calculate, at each trading time ti , the joint posterior of (θ,
marginal posteriors, their Bayes estimates and their standard errors (SE),
respectively. The recursive algorithm and the software are fast enough to
generate real-time Bayes estimates. We test the software extensively and
verify it on Monte Carlo data, where we know the true parameters.
Figure 5 has four plots to display how the trade-by-trade Bayes esti-
mates and the two standard errors (SE) bounds evolve in comparison with
the true values for µ, σ, κ and ρ, respectively. The figure clearly shows that
the estimates of µ, σ, κ and ρ converge to their true values, the two-SE
bounds shrink, and the true values are within the two-SE bounds. Simi-
larly, the Bayes estimates of X(t) are close to their true values, which are
always within two-SE bounds.Their final Bayes estimates, SEs, and true
values are presented in Table 1. The true values are close to the Bayes
estimates and all within two SE bounds.
5. An Empirical Study of 5-year Treasury Note

Our data are provided by GovPx, Inc, an organization owned by primary
dealers and three inter-dealer brokers to collect and distribute real-time
quotes and transaction data of U.S. Treasury Securities. GovPx, Inc. records
Fig. 5.
−7 mu
x 10
1.5
0.5
0
0 500 1000 1500 2000
−5 sigma
x 10
1.5
1.4
1.3
1.2
0 500 1000 1500 2000
rho
0.03
0.025
0.02
0.015
0 500 1000 1500 2000
−6 kapa
x 10
−1
−2
−3
−4
−5
0 500 1000 1500 2000
Bayesian Estimates Upper 2−SE Bound Lower 2−SE Bound True value
Table 1. Bayes Estimates for an Monte Carlo Data
Parameter True Value Bayes Estimate St. Error

µ 6.32E-8 5.63E-8 1.09E-8
σ 1.33E-5 1.32E-5 3.25E-7
κ -2.71E-6 -3.04E-6 3.94E-7
ρ 2.280E-2 2.19E-2 2.23E-3
The value of parameters are for per second.
all trades of Treasury Securities by the three major interdealer brokers ex-
cept one major interdealer broker, Cantor Fitzgerald. Thus its data count
for a significant portion of the primary-dealer activity. We exam trades
of the active, “on-the-run”, 5-year U.S. Treasury notes issued on Novem-
ber 15, 2000 and for the period 11/15-12/29, 2000. During this time period,
GovPx-brokered trades in Treasury notes count for roughly half of the trad-
ing volume in notes on a global basis. A big bulk of trading activity in the
treasury market is in the active instruments, so-called “on-the run” instru-
ment which is the most-recently auctioned security of that type. Secondary
trading in U.S. Treasury securities takes place in an over-the-counter mar-

ket. Although the global market is open around-the-clock during the week,
95 percent of the trading occurs during New York trading hours, from
around 7:30 A.M. to 5:00 P.M. Therefore, we restrict our series to the
trades that are executed between 7:30 AM to 5:00 PM Eastern Time. The
basic summary statistics can be found in Table 2.
Table 2. Summary Statistics for 5-year Treasury Note 11/15 - 12/29, 2000
Size Mean Median St. Dev. Skewness Kurtosis

5-year note 2,220 101.83% 101.84 1.03% 9.65E-3 1.745
The relative frequencies for odd 1/128, odd 1/64 and odd 1/32 or coarser
are 18.24%, 37.52%, and 44.23%. Method of relative frequencies produces
the estimates for the clustering parameters: α = .2505 and β = .3847.
The daily pattern of treasury trading has been studied in details Flem-
ing.9 To model the daily duration pattern for the transactions of U.S. trea-
sury, we follow the same approach as in Engle and Russell.6 Figure 5 shows
the nonparametric kernel estimation (run by supsmu function in Matlab) of
the expected duration conditional on the time of the day. Consistent with
previous studies, trading is most active between 8:30 a.m and 9:30 which
is partly a result of the important macroeconomic news release time and
the opening of U.S. Treasury futures market. The slight shorter durations
around 2:30 p.m. and 3:00 p.m. coincides with the closing time of U.S.
Treasury futures market. To capture this diurnal pattern, we use a cubic
spline (run by spfit function in Matlab) to approximate the daily factor.
Nodes are set on 7:30 a.m., 8:30 a.m., 9:00 a.m. and each hour after 9:00
a.m. until 5:00 p.m. when the trading activity falls.
After the diurnal adjustment for duration (namely, each ∆ti divided by
a positive number produced by the cubic spline fit according to the time
of day), we fit an EACD(1, 1) model using maximum likelihood approach.
The parameter estimates are given in Table 3. All parameters are signif-
icant. The sum of α1 and β1 is 0.9377, indicating strong persistence in
duration. We do diagnostic check on EACD(1,1) model. If the model is ap-
propriate, the residues, or the standardized durations, are i.i.d. exponential
distributed with mean one. Higher order moments should be independent
also. The Ljung-Box Q-Statistics for residue and residue-squared suggest
that the model does a very good job of accounting for the intertemporal
dependence in durations. Next, we consider the exponential specification of
Fig. 6.
Nonparametric estimate of daily pattern for transaction durations(5Year2ndInstrument)
600
550
500
Duration(seconds)
450
400
350
300
250
7:30AM 8:30AM9AM 10AM 11AM 12PM 1PM 2PM 3PM 4PM 5PM
Time
residue. For the null hypothesis that the residue is exponential, Engle and
Russell6 (page 1144) provided a simple test, whose limiting distribution is
standard normal and the test statistic is reported in the last line of Table
3. Even though the exponential assumption is rejected, the test value is
much smaller than those in Engle and Russell.6 This also suggests other
distributions such as Weibull or generalized Gamma distributions may be
considered in the future research. Even for other ACD models, since they
are FtY,V -predictable, the Bayes estimates of (µ, σ, κ, ρ) presented in Table
4 are not changed and these estimates are “model-free” of the assumptions
of duration.
With the estimates of the non-clustering parameters, we apply the tested
Fortran program to the transaction data of 5-year bond and obtain Bayes
estimates for the extended FM model for two cases. One is GBM case and
the other is GBM with initiation dummy in volatility. The Bayes estimates
of both cases are presented in Table 4. All the estimates are statistically
significant. Note that the estimates of µ and ρ changes quite significantly,
suggesting the significant impact of including the initiation dummy. More-
over, the sign κ is negative, implying inventory theory is the main impact
in the bond trading.
Table 3. Parameter Estimates for EACD(1,1) Model
Estimates Std Error T-Stat Significance

α0 0.06225 0.01177 5.290 1.2E-7
α1 0.06542 7.56E-3 8.657 0
β1 0.8723 0.01672 52.17 0
Ljung-Box Q-Statistics of Adjusted Durations
Q(10-0) 136.72 0
Q(20-0) 168.28 0
Ljung-Box Q-Statistics of Residue
Q(10-0) 5.4943 0.8558
Q(20-0) 16.276 0.6993
Ljung-Box Q-Statistics of Residue- Squared
Q(10-0) 3.607 0.9633
Q(20-0) 13.607 0.8499
Test Statistic 4.2134 1.258E-5
Table 4. Annualized Bayes Estimates for 5-year Bond, 11/15-12/29, 2000
Model µ σ κ ρ
GBM only 25.10% 3.41% 0 0.128
(κ ≡ 0) (10.01%) (0.06 %) (0) (0.0147)
Extended GBM 54.50% 3.90% -0.80% 0.0228
(κ 6= 0) (10.63%) (0.11%) (0.14%) (0.0085)
Standard errors are in parentheses.
6. Conclusion
This paper reviews recent development of a rich class of filtering models
with counting process observations for the micromovement of asset price
with observable factors and the related Bayes estimation via filtering. A
specific extended FM model built upon GBM and with a buyer-seller initi-
ation dummy in volatility is constructed for the 5-year Treasury note trans-
action price data. Employing the developed filtering techniques, we perform
Bayes estimation via filtering for the model. We find that the buyer-seller
initiation dummy is statistically significantly negative. This supports the
inventory theory for the bond trading.
This model can be further extended in different directions. For example,
we can study whether other observable factors have impacts on volatility.
Especially, in Hu, Kuipers and Zeng,16 we add another macro-economic
news dummy in the volatility and we find that macro-economic news has
even higher impacts on the bond price volatility than the buyer-seller initi-
ation dummy. With these two examples, we believe the method developed
should have many financial applications in empirical market microstructure

theory.
Appendix A. More on Clustering Noise

To formulate the biasing rule, we first define a classifying function r(·),

 2 if the fractional part of y is odd 1/128
r(y) = 1 if the fractional part of y is odd 1/64 . (A.1)
0 if the fractional part of y is odd 1/32 or coarser

The biasing rules specify the transition probabilities from y 0 to y, p(y|y 0 ).

Then, p(y|x), the transition probability can be computed through p(y|x) =
0 0 0 0 1
P
y 0 p(y|y )p(y |x) where p(y |x) = P {V = y − R[x, 8 ]}. Suppose D =
1
8|y−R[x, 8 ]|. Then, p(y|x) can be calculated as, for example, when r(y) = 2,

 (1 − ρ)(1 + αρ) if r(y) = 2 and D = 0
p(y|x) = 21 (1 − ρ)[ρ + α(2 + ρ2 )] if r(y) = 2 and D = 1 . (A.2)
1 D−1 2
2 (1 − ρ)ρ [ρ + α(1 + ρ )] if r(y) = 2 and D ≥ 2
Acknowledgments
Part of the work was done when Yong Zeng was on sabbatical, visited and
taught a special course on Statistical Analysis of Ultra-high Frequency Fi-
nancial Data - An Overview and A New Filtering Approach in Department
of Operations Research and Financial Engineering (ORFE) in Princeton
University in Spring semester of 2007. He would like to thank Savas Dayanik
for the invitation and thank ORFE Department of Princeton for the hospi-
tality. He also gratefully acknowledge the support of the National Science
Foundation under grant DMS-0604722.
References
1. Ait-Sahalia, Y., P. A. Mykland and L. Zhang, How Often to Sample a
Continuous-Time Process in the Presence of Market Microstructure Noise,
Review of Financial Studies, 18 (2005), 351–416.
2. Bandi, F. M. and J. R. Russell, Separating microstructure noise from volatil-
ity, Journal of Financial Economics, 79 (2006), 655-692.
3. Bremaud, P., 1981. Point Processes and Queues:Martingale Dynamics.
Springer-Verlag, New York.
4. Cvitanic, J., R. Liptser and B. Rozovskii, A filtering approach to tracking
volatility from prices observed at random times. Annals of Applied Probabil-
ity, 16 (2006), 1633–1652.
5. Engle, R. F., The econometrics of ultra-high-frequency data, Econometrica,

68 (2000), 1–22.
6. Engle, R. F. Engle and J. R. Russell, Autoregressive Conditional Duration:
A New Model for Irregularly Spaced Transaction Data, Econometrica, 66
(2000), 1127–1162.
7. Ethier, S. N. and T. G. Kurtz, 1986. Markov processes : Characterization and
convergence. Wiley, New York.
8. Fan, J. and Y. Wang, Multi-scale jump and volatility analysis for high-
Frequency financial data, Journal of American Statistical Association, 102
(2007), 1349-1362.
9. Fleming, M. J., The Round-the-Clock Market for U.S. Treasury Securities,
Economic Policy Review - Federal Reserve Bank of New York, 3 (1997), 9-32.
10. Fleming, M. J. and J. V. Rosenberg, How do Treasury dealers manage their
positions? it SSRN Working Paper #972367, Federal Reserve Bank of New
York, 32 pp.
11. Frey, R., Risk-minimization with incomplete information in a model for high-
frequency data, Mathematical Finance, 10 (2000), 215-225.
12. Frey, R. and W. J. Runggaldier, A nonlinear filtering approach to volatility
estimation with a view towards high frequency data, International Journal
of Theoretical and Applied Finance 4 (2001), 199-210.
13. Hasbrouck, J., 1996. Modeling market microstructure time series, in G. Mad-
dala and C. Rao, eds, ‘Handbook of Statistics’, Vol. 14, North-Holland, Am-
sterdam, pp. 647-692.
14. Hasbrouck, J., Stalking the ”efficient price” in market microstructure speci-
fications: an overview, Journal of Financial Markets 5 (2002), 329-339.
15. Harris, L., Stock price clustering and discreteness, Review of Financial Stud-
ies, 4 (1991), 389-415.
16. Hu, X, D. Kuipers and Y. Zeng, Econometric analysis via filtering for ultra-
high frequency data, Working paper (2010). University of Missouri at Kansas
City.
17. Kouritzin, M. and Y. Zeng, Bayesian model selection via filtering for a class of
micro-movement models of asset price, International Journal of Theoretical
and Applied Finance, 8 (1991), 97-121.
18. Lee, C. M. C. and M. J. Ready, Inferring Trade Direction from Intraday
Data, Journal of Finance, 46 (1991), 733-746.
19. Lee K. and Y. Zeng, Risk minimization for a filtering micromovement model
of asset price, Applied Mathematical Finance, 17 (2010), 177–199.
20. Li, Y. and P. A. Mykland, Are volatility estimators bobust with respect to
modeling assumptions? Bernoulli, 13 (2007), 601-622.
21. O’Hara, M., 1997. Market microstructure theory. Blackwell Publishing Inc.,
Malden, MA.
22. Protter, P. 2003. Stochastic Integration and Differential Equations, Springer-
Verlag, Berlin, 2nd Edition.
23. Stroock, D. W. and S. R. S Varadhan, 1979. Multidimensional Diffusion
Processes, Springer-Verlag, Berlin.
24. Xiong, J. and Y. Zeng, Mean-variance portfolio selection for partially-
observed point process. Working paper (2010). University of Tennessee.

25. Zeng, Y., A partialy observed model for micromovement of asset prices with
Bayes estimation via filtering, Mathematical Finance, 13 (2000), 411–444.
26. Zeng, Y., Bayesian inference via filtering for a class of counting processes:
Application to the micromovement of asset price, Statistical Inference for
Stochastic Processes, 8 (2005), 331 – 354.
27. Zhang, L., Mykland, P. A. and Aı̈t-Sahalia, Y., A tale of two time scales:
Determining integrated volatility with noisy high frequency data, Journal of
the American Statistical Association, 100 (2005), 1394–1411.
May 27, 2011 13:45 WSPC - Proceedings Trim Size: 9in x 6in 07-koh
145
Jump Bond Markets Some Steps towards General Models in

Applications to Hedging and Utility Problems
Michael Kohlmann
Department of Mathematics and Statistics
University of Konstanz
D-78457, Konstanz, Germany
Email: michael.kohlmann@uni-konstanz.de
Dewen Xiong
Shanghai Jiaotong University
Shanghai 200240, P.R.China
Email: xiongdewen@sjtu.edu.cn
In finance, a bond is a debt security in which the authorized issuer owes the
holder a debt and, depending on the terms of the bond, is obliged to pay interest
(the coupon) and/or to repay the principal a a later date, termed maturity.
A bond is a formal contract to repay borrowed money with interest at fixed
intervals.
Thus a bond is like a loan: The issuer is the borrower, the holder is the
lender, and the coupon is the interest. Bonds provide the borrower with external
funds to finance long-term investments, or, in the case of government bonds,
to finance current expenditure. Bonds must be repaid a fixed intervals over a
period of time.
Bonds and stocks are both securities, but the major difference between
the two is that stockholders ave an equity stake in the company (i.e. they are
owners), whereas bondholders have a creditor stake in the company (i.e. they
are lenders). Another difference is that bonds have a defined term or maturity
after which the bond is redeemed, whereas may be outstanding indefinitely.
The simplest case of a bond is a bank account with fixed interest rate rt :
dP (t, T ) = rt P (t, T )dt, P (T, T ) = 1.
A short look at the chart of a banks interest rates over some years however
shows that the interest rate is by no means fixed so that we should assume that
it is a random variable. This leads to the well known classical bond models (in
alphabetical order of the authors) by Black-Derman, Chen, Cox-Ingersoll-Ross
(CIR), Heath-Jarrow-Morton (HJM), Ho-Lee, Hull-White, Vasicek (V), among
many others, where those with an abbreviation after the names are the best
known to my observation.
However, while models for stock markets are meanwhile quite satisfactory,
the models for bonds/interest rates have diverse deficiencies, so that reality
146 M. Kohlmann and D. Xiong
is not really well described. Also the models appear mathematically more dif-
ficult and technical than the classical stock market models: While researchers
and practitioners are concerned with the fine-tuning in stock models, the basics
on a general model for bond markets are still discussed without a commonly
agreed result.
The first attempt to describe a general model is found in the seminal paper
by Björk et al(1997)7 and we are often referring to this article.
Here we will discuss two mainstreams of problems, namely mean variance hedg-
ing and utility optimization (exponential utility indifference) in a general jump
bond market. The purpose of this paper then is to introduce some new tech-
niques, especially techniques from the theory of backward stochastic differential
equations (BSDEs) in mean variance hedging, and to contribute to a general
model.
In the first part we will consider a model based on n maturities to apply
recent results from MVH in jump stock markets. Carmona et al(2004)10 im-
pressively describe the shortcomings of the models based on a finite number of
maturities. To make things short: In the corresponding continuous HJM model
the market is infinite dimensional with only a finite number of random sources
so that this market always is complete which is contrary to the observations.
Further shortcomings caused by the infiniteness of the market are described in
Cont (2004)17 .
There are several attempts to overcome these difficulties. Carmona et
al(2004)10 introduce an infinite dimensional Brownian motion and so use Malli-
avin calculus methods to treat hedging problems with hedging strategies de-
rived from a Clark-Ocone formula. Baran et al.2 consider generalized strate-
gies in an infinite dimensional HJM-model. A similar approach is taken by
DeDonno18 . However these generalized strategies are not very useful to solve
hedging problems more explicitly. In the second part of this paper we will here
propose and extend an infinitely dimensional market where we consider mea-
sure valued strategies. Of course also these strategies have certain drawbacks
when we come to the economical interpretation. And the main drawback is the
fact that we always have to work in martingale markets to consider (M, Q0 )-
normalized martingales as approximate wealth processes. For many problems
this is sufficient but e.g. to describe superhedging we needed the notion of
semimartingale. This, however, still is a long-standing open problem already de-
scribed by Schwartz(1994)43 : Il n’est pas facile de savoir exactement de qu’on
doit appeler une semimartingale valeurs dans un espace vectoriel topologique.
So, the contents of this paper should be seen as one perhaps promising looking
step towards a general model for which we then consider exponential hedging
problems.
The paper is organized as follows: In the first part we construct a market of
bonds with jumps driven by a general marked point process as well as by an Rn -
valued Wiener process based on Björk et al(1997)7 , in which there exists at least
one equivalent martingale measure Q0 . Then we consider the mean-variance
hedging of a contingent claim H ∈ L2 (FT0 ) w.r.t. self-financing portfolios
based on the given maturities T1 , · · · , Tn with T0 < T1 < · · · < Tn ≤ T ∗ .
We introduce the concept of variance-optimal martingale (VOM) and describe
the VOM by a backward semimartingale equation (BSE). By making use of
the concept of E ∗ -martingales introduced by Choulli et al.(1998)13 , we obtain
another BSE which has a unique solution. We derive an explicit solution of
the optimal strategy and the optimal cost of the mean-variance hedging by the
solutions of these two BSEs.
Jump Bond Markets Some Steps towards General Models 147
In the second section we consider the optimal exponential utility in a

bond market with jumps basing on a model similar to Björk, Kabanov and
Runggaldier(1997)7 which is arbitrage-free. Similar to the normalized integral
with respect to the cylindrical martingale first introduced in Mikulevicius and
Rozovskii(1998)39 , we introduce the (M, Q0 )-normalized martingale and local
(M, Q0 )-normalized martingale. For a given maturity T0 ∈ [0, T ∗ ], we describe
the minimal entropy martingale (MEM) based on [T0 , T ∗ ] by a backward semi-
martingale equation (BSE) w.r.t. the (M, Q0 )-normalized martingale. Then we
give an explicit form of the optimal approximate wealth to the optimal exp-
utility problem by making use of the solution of the BSE. Finally, we describe
the dynamics of the exp utility indifference valuation of a bounded contingent
claim H ∈ L∞ (FT0 ) by another BSE under the minimal entropy martingale
measure in the incomplete market.
The present paper strongly relies on unpublished works 45–48 of the au-
thors and a seminar talk of the first author at the Nomura Institute of Oxford
University in October 2009. Full proofs of results left out in this report are
found in the cited preprints.
Keywords: Bond market with jumps; BSE; variance optimal martingale

(VOM); E ∗ -martingale; mean-variance hedging (MVH); minimal entropy mar-
tingale (MEM); (M, Q0 )-normalized martingale; exp utility indifference valua-
tion (exp-UIV).
Mathematics Subject Classification. 90A09, 60H30, 60G44.
1. Preliminaries: The Basic Model

We consider a financial market model on a stochastic basis (Ω, F , F, P ),
where F = {Ft ; t ∈ [0, T ∗ ]} is a filtration satisfying the usual conditions
with FT ∗ = F . The basis is assumed to carry an n-dimensional Brownian
motion W = (W1 , · · · , Wn )0 as well as an integer-valued random measure
µ(du, dy) on R+ × R with compensator ν(du, dy) = λ(u, dy)du.
Assumption 1.1. The filtration F = {Ft ; t ∈ [0, T ∗ ]} is the natural filtra-

tion generated by W and µ, i.e.,

Ft = σ W (s), µ([0, s] × A), B; 0 ≤ s ≤ t, A ∈ B, B ∈ N
where N is the collection of P -null sets from F .
For example from remark 3.2 of Björk et al (1997)7 the stochastic basis
has the following predictable representation property: any local martingale
M is of the form
Z t Z tZ
Mt = M0 + fu0 dW (u) + ψ(u, y)(µ(du, dy) − ν(du, dy))
0 0 R
where f is an Rn -valued predictable process and ψ is a P-measurable

e func-
tion such that
Z T Z TZ
|fu |2 du < ∞, |ψ(u, y)|λ(u, dy)du < ∞.
0 0 R
In this paper, we let the 4-tuple (σ, δ, φ, ϕ) be given such that

(i) σ(u, T ) is a bounded Rn -valued predictable process on [0, T ], and
δ(u, y, T ) is a bounded P-measurable
e function on [0, T ] × E for ev-
ery fixed T ∈ [0, T ∗ ]; we assume that both σ(u, T ) and δ(u, y, T ) are
continuously differentiable in the T -variable, and we set
Z T
D(u, y, T ) = − δ(u, y, l)dl;
u
(ii) φ is a bounded R -valued predictable process on [0, T ∗ ];

n
(iii) ϕ is a bounded strictly positive P-measurable

e function on [0, T ∗ ] × E.
We introduce 16
Z T
Ψ(u, T ; σ, δ, φ, ϕ) : = σ(u, T )0 σ(u, s)ds − σ(u, T )0 φ(u)
u
Z
− δ(u, y, T )eD(u,y,T )ϕ(u, y)λ(u, dy),(1)
E
we let the forward rate dynamics be given by 15
Z t
llf (t, T ) = f (0, T ) + Ψ(u, T ; σ, δ, φ, ϕ)du (2)
0
Z t Z tZ
0
+ σ(u, T ) dW (u) + δ(u, y, T )µ(du, dy)
0 0 E
7
From Proposition 2.2 of Björk et al(1997) the short rate rt := f (t, t) is
then given by
Z t Z t Z tZ
rt = r0 + αu du + σ(u, u)0 dW (u) + δ(u, x, u)µ(du, dy),
0 0 0 E
where
Z
0
αu = fT (u, u) − σ(u, u) φ(u) − δ(u, y, u)eD(u,y,u) ϕ(u, y)λ(u, dy)
E
and

∂f (u, T )
fT (u, u) := .
∂T T =u
Since
Z T
A(u, T ) := − Ψ(u, l; σ, δ, φ, ϕ)dl
Z uT
l
Z Z T
σ(u, l)0 σ(u, l)0 φ(u)dl

=− σ(u, s)ds dl +
u u u
T
Z Z
δ(u, y, l)eD(u,y,l) dl ϕ(u, y)λ(u, dy)

+
E u Z
1 D(u,y,T )
= − kS(u, T )k2 − S(u, T )0 φ(u) −

e − 1 ϕ(u, y)λ(u, dy),
2 E
where S(u, T ) = (S1 (u, T ), · · · , Sn (u, T ))0 and where Si (u, T ) :=

Z T
− σi (u, l)dl, one sees from Björk, Kabanov and Runggaldier(1997)7 that
u RT
the bond price process p(t, T ) := e− t
f (t,s)ds
can be rewritten as
Zt 1
p(u−, T ) ru + A(u, T ) + kS(u, T )k2 du

p(t, T ) = p(0, T ) +
Z t 0 Z tZ2
p(u−, T ) eD(u,y,T ) − 1 µ(du, dy)

+ p(u−, T )S(u, T )0 dW (u) +
0 Z t Z t 0 E
= p(0, T ) + p(u−, T )ru du + p(u−, T )S(u, T )0 dW̃ (u)
Z tZ 0 0
p(u−, T ) eD(u,y,T ) − 1 µ(du, dy) − ϕ(u, y)λ(u, dy)du ,

+
0 E
Z t
where W̃ (t) := W (t) − φ(u)du. As φ is a bounded Rn -valued process
0
and ϕ a bounded strictly positive P-measurable
e function, ϕ(u, y) − 1 > −1,
we have that
Z · Z ·Z
Zt0 = E φ(u)0 dW (u) +

ϕ(u, y) − 1 µ(du, dy) − ν(du, dy)
Z0 t 0 E Z tZ t
0 0 0
=1+ Zu− φ(u) dW (u) + Zu− ϕ(u, y) − 1 µ(du, dy) − ν(du, dy) ,
0 0 E
is a strictly positive uniformly integrable martingale on [0, T ∗]. So we can

define a new probability Q0 ∼ P by
dQ0

:= ZT0 ∗ .
dP F
It follows from Girsanov’s theorem that W̃ is a Brownian motion under Q0

and
ν̃(du, dy) := ϕ(u, y)λ(u, dy)du

is the compensator of µ(du, dy) under Q0 , thus the dynamics of the bond
price under Q0 are given by
Z · Z ·
p(t, T ) = p(0, T )E ru du + S(u, T )0 dW̃ (u)
0 Z Z 0
· D(u,y,T )

+ e − 1 (µ(du, dy) − ν̃(du, dy)) .
0 E t
Rt
− ru du
The discounted bond price p̄(t, T ) = e p(t, T ) is then given by
0
Z ·
p̄(t, T ) = p̄(0, T )E S(u, T )0 dW̃ (u)
0 Z Z
· D(u,y,T )

+ e − 1 (µ(du, dy) − ν̃(du, dy)) ,
0 E t
(3)
which is a uniformly integrable Q0 -martingale, by the boundedness of S
and D.
Definition. (also see 7) We say that a probability Q on (Ω, F ) is a martin-

gale measure if Q ∼ P and if the discounted bond price {p̄(t, T ); t ∈ [0, T ∗ ]}
is a local martingale under Q for each T ∈ [0, T ∗ ].
Remark. As Q0 is a martingale measure the market is arbitrage free (also

see Proposition 3.9 of Björk et al(1997)7). From 7, we adopt the following
lemma
Lemma 1.1. Let Q be a martingale measure and Z be the density process
of Q with respect to Q0 , then Z can be represented in the following form
Z · Z ·Z
Zt = E z1 (u)0 dW̃ (u) + z2 (u, y)(µ(du, dy) − ν̃(du, dy)) (4)
0 0 E t
n
where z1 is an R -valued predictable process and z2 is a P-measurable
e func-
tion with z2 (u, y) > −1 such that for all u ∈ [0, T ] and T ∈ [0, T ∗ ] the
following structure condition holds
Z
z1 (u)0 S(u, T ) +
D(u,y,T )
e − 1 z2 (u, y)ϕ(u, y)λ(u, dy) = 0. (5)
E
We have:
Corollary 1.1. When σ(u, T ) is independent on the maturity T , i.e.,
σ(u, T ) ≡ σ(u), where {σ(u); u ∈ [0, T ∗]} is a bounded Rn -valued predictable
process, and δ(u, y, T ) is of the following form
δ̃(u, y)
δ(u, y, T ) = −
1 + δ̃(u, y)(T1 − u)
1
for a bounded P-measurable
e function δ̃(u, y) with δ̃(u, y) > − ∗ , then
T
(z1 , z2 ) in the density process must satisfy
Z
−z1 (u)0 σ(u) + δ̃(u, y)z2 (u, y)ϕ(u, y)λ(u, dy) = 0. (6)
E
Obviously the market is incomplete.
2. MVH in Jump Bond Markets

The mean-variance hedging in stock jump markets has already been dis-
cussed in e.g. Arai(2005)1 , Černý-Kallsen(2007)11 and Kohlmann-Xiong-
Ye(2007)31,35 . Arai(2005)1 adapted the Gouriéroux-Laurent-Pham (1998)25
(GLP-)approach to the discontinuous case mainly under the assumption
that the VOMM exists as a probability measure. A decomposition of H on
S is derived by the same sort of argument as in the alternative approach in
Section 4 of Rheinländer and Schweizer (1997)42 , so that the decomposition
of H on S, in general, is not an orthogonal one, that is, not a GKW one.
Another work about mean-variance hedging in a general semimartingale
setting is Černý-Kallsen(2007)11. They introduce a new measure (called
’opportunity-neutral measure’) P ? such that the minimal martingale mea-
sure relative to P ? coincides with the variance-optimal martingale measure
relative to the original probability measure P . Then a so-called opportunity
process and an adjustment process are introduced which give the densities
of the VOMM and P ? . Basing on these tools methods are developed to
carry over projection methods -inspired by the GKW- or GPL-method in
simpler settings- to derive a formula for the optimal hedge in feedback form
(Lemma 4.9). However they avoid to use control theoretical tools -although
the new tools have their well understood counterparts there- and do not
show the connection to the meanwhile well developed BSDE approach to
the MVH-problem. In a different approach, Kohlmann-Xiong-Ye(2007)31,35
relate the variance optimal martingale measure (VOMM) to a backward
semimartingale equation (BSE) and show that the (VOMM) is equivalent
to the original measure P if and only if the BSE has a solution. For a general
contingent claim, they derive an explicit solution of the optimal strategy
and the optimal cost of the mean-variance hedging by means of another
backward semimartingale equation and an appropriate predictable process
δ.
In this part of the paper, we continue the work of Kohlmann-Xiong-
Ye(2007)31,35 and discuss the mean-variance hedging in a market of bonds
with jumps, in which the variance optimal martingale measure may be
negative, from the point of view of stochastic control. Based on some

adapted facts about the adjustment process (see Černý-Kallsen(2007)11 or
Schweizer(1996)44), we introduce the variance-optimal martingale M ∗ , a
Q0 -local martingale whose jumps may even be equal to −1. Then we derive
a nonlinear backward semimartingale equation (see (6)), and give an ex-
plicit form of the variance-optimal martingale M ∗ by the bounded solution
of the BSE (6). In section (2.2), we introduce the E ∗ -martingale, which is
similar to the concept of E-martingales introduced in Choulli-Krawczyk-
Stricker(1998)13. By applying the representation of an E ∗ -martingale, we
derive another linear BSE (see (7)) whose parameters depend on the BSE
(6), and we derive an explicit solution of the optimal strategy and the op-
timal cost of the mean-variance hedging by means of this solution of the
BSE (7). In section (2.3), we consider a special bond market where σ in
(15) is independent of the maturity T and δ is of the following form
δ̃(u, y)
δ(u, y, T ) = −
1 + δ̃(u, y)(T1 − u)
1
e function δ̃(u, y) with δ̃(u, y) > − ∗ . In this
T
special case we derive the optimal strategy in a very explicit form.
2.1. Hedging on the Basis of n Maturity Times and the

Variance Optimal Martingale Measure
We consider a market as described above in section 1.
Now fix a T0 ∈ (0, T ∗ ) and let T1 , · · · , Tn be n maturities with T0 <
Ti < T ∗ for each i and T1 < T2 < · · · < Tn . We are mainly concerned
with the question whether a contingent claim given by H ∈ L2 (FT0 ) can
be replicated or hedged by a self-financing portfolio based on
p̄(t, T1 ), . . . , p̄(t, Tn ),
which is defined in the following way:
Definition. A self-financing portfolio based on the maturities

T1 , · · · , Tn is an Rn -valued predictable process π = (π1 , · · · , πn )0 such that
the corresponding discounted value process
Z tXn n
Z tZ X
Vtx,π = x + πj (u)S(u; Tj )0 dW̃ (u) πj (u){eD(u,y,Tj ) − 1}
0 j=1 0 E j=1

µ(du, dy) − ν̃(du, dy) ,
(1)
t ∈ [0, T0 ], is a local Q0 -martingale with VTx,π ∈ L2 (FT0 , P ). The family of

all self-financing portfolios based on the maturities T1 , · · · , Tn is denoted
by A(T1 , · · · , Tn ).
Of course in general, a contingent claim H ∈ L2 (FT0 , P ) can not be fully

replicated by a self-financing portfolio based on the maturities T1 , · · · , Tn ,
thus we discuss the following minimization question
h 2 i
min E H − VTx,π 0
,
π∈A(T1 ,··· ,Tn )
and give an explicit solution of the optimal strategy by BSDEs.

Let
 
Z is a local Q0 -martingale on [0, T ] with Z = 1 and


 0 0 


h i
Z2s (T1 , · · · , Tn ) := Z EQ0 ZT00 (ZT0 )2 < ∞ such that Zt p̄(t, Ti ) is a local Q0 - ,

 
 martingale on [0, T ] for each i 

0

Z2exp (T1 , · · · , Tn ) := Z; Z ∈ Z2s (T1 , · · · , Tn ), Z is a stochastic exponential ,

Z2e (T1 , · · · , Tn ) := Z; Z ∈ Z2s (T1 , · · · , Tn ), Z is strictly positive ,
Obviously we have Z2e (T1 , · · · , Tn ) ⊂ Z2exp (T1 , · · · , Tn ) ⊂ Z2s (T1 , · · · , Tn ).

dQ
For any Z ∈ Z2e (T1 , · · · , Tn ), one can define a measure Q by :=
dQ0 FT0
ZT0 , and then Q is an equivalent martingale measure for each p̄(t, Ti ), i =
1, · · · , n.
Remark. From Jacod(1979)26 the following is well known: for any Z ∈
Z2exp (T1 , · · · , Tn ), let τ := inf{t : Zt = 0} then Zt− 6= 0 a.s. for 0 ≤ t ≤ τ ,
and Zt = 0 a.s. for τ ≤ t ≤ T0 . We note that Z may be negative.
We define for any t, u ∈ [0, T0 ] and y ∈ E

 
S1 (t, T1 ) · · · Sn (t, T1 )
S(t) :=  · · · ··· ···  = (Sj (t, Ti ))n×n ,
S1 (t, Tn ) · · · Sn (t, Tn ) n×n
0
Υ(u, y) := eD(u,y,T1 ) − 1, · · · , eD(u,y,Tn ) − 1 ,
0
P̄ (t) := (p̄(t, T1 ), · · · , p̄(t, Tn )) ,
and, similar to Assumption (H2) in Biagini(2001)4, we assume in this sec-
tion that
(H) S(t) is invertible for all t ∈ [0, T0 ].
Lemma 2.1. Z ∈ Z2exp if and only if Z can be represented as

Z t Z tZ
Zt = 1 + Zu− z1 (u)0 dW̃ (u) + Zu− z2 (u, y)(µ(du, dy) − ν̃(du, dy))
0 0 E
where z1 is a Rn -valued predictable process and z2 is a P-measurable

e func-
tion such that for all u ∈ [0, T0 ]
Z
S(u)z1 (u) + Υ(u, y)z2 (u, y)ϕ(u, y)λ(u, dy) = 0. (2)
E
We denote by TT0 the family of stopping times τ with τ ≤ T0 . For any

given stopping time τ ∈ TT0 and (z1 , z2 ) satisfying (2), let
Z ·
E τ (z1 , z2 ) : = E I]]τ,T0 ]] (u)z1 (u)0 dW̃ (u)
0 Z ·Z
+ I]]τ,T0 ]] (u)z2 (u, y)(µ(du, dy) − ν̃(du, dy)) ,
0 E
2,τ
and we write Zexp for the family of (z1 , z2 ) satisfying (2) with
2
E Zτ,T0 E τ (z1 , z2 )T0
0
< ∞. We consider the following kind of minimiza-
tion problem
h
0
2 i
Ψ(τ ) := ess inf 2,τ E Zτ,T0
E τ(z1 , z2 )T0 Fτ . (3)
(z1 ,z2 )∈Zexp
The following lemma is adapted from 11,34,35
Lemma 2.2. There exists an Rn -valued predictable process ã =

(ã1 , · · · , ãn )0 (called adjustment process) on [0, T0 ] such that for any
τ ∈ TT0 , E −ã0 I]]τ,T0 ]] · P̄ T ∈ L2 (P ) and such that

0
0
0 τ
(i) E −ã I]]τ,T0 ]] · P̄ s Zτ,s E (z1 , z2 )s ; s ∈ [τ, T0 ] is a uni-
formly integrable martingale for any τ ∈ TT0 and (z1 , z2 ) ∈
2,τ
Zexp ;
h i
(ii) E E −ã0 I]]τ,T0 ]] · P̄ T Fσ

0

0
2
= E E −ã I]]τ,T0 ]] · P̄ T Fσ ∈ (0, 1], a.s., for any
0
stopping times σ ≤ τ ;
(iii) for any π ∈ A(T1 , · · · , Tn )

E VTx,π − Vτx,π E −ã0 I]]τ,T0 ]] · P̄ T0 = 0.

0
From Lemma 2.2, it appears to be natural to assume that

n o
(τ )
Assumption 2.1. For any τ ∈ TT0 , Mt ; t ∈ [τ, T0 ] is a stochastic
exponential, where

E E −ã0 I]]τ,T0 ]] · P̄ T0 Ft

(τ )
Mt := .
E E −ã0 I]]τ,T0 ]] · P̄ T Fτ
0
(τ )
Under Assumption 2.1, {Mt (p̄(t, Ti ) − p(τ, Ti )); t ∈ [τ,
T0 ]} is a local
(τ )
martingale, thus Z10 Mt (p̄(t, Ti ) − p(τ, Ti )); t ∈ [τ, T0 ] is a local Q0 -

t
(τ ) (τ )
martingale, therefore there exists a pair (z1 , z2 ) depending on τ satisfy-
ing (2) such that

(τ ) 0 (τ ) (τ )
Mt = Zτ,t E τ z1 , z2 , a.s., for t ∈ [τ, T0 ], (4)
t
and
" #
2
0 (τ ) (τ )
Ψ(τ ) = E Zτ,T Eτ z1 , z2

0
Fτ
T0
(τ ) 2
= E MT0 Fτ
1
= h i.
E E −ã0 I]]τ,T0 ]] · P̄ T0
Fτ
Lemma 2.3. Under Assumption 2.1, there exists a pair (z1∗ , z2∗ ) satisfying
(2) such that for all τ ∈ TT0 , (z1∗ , z2∗ ) ∈ Zexp2,τ
and
h
0
2 i
Ψ(τ ) = EQ0 Zτ,T 0
E τ(z1∗ , z2∗ )T0 Fτ , Q0 -a.s.

(τ ) (τ )
Proof. We only need to show that z1 , z2 in (4) is independent of τ ,
i.e., for any τ1 , τ2 ∈ TT0 with τ1 < τ2 , we need to show that z1τ1 I]]τ2 ,T0 ]] =
z1τ2 I]]τ2 ,T0 ]] and z2τ1 I]]τ2 ,T0 ]] = z2τ2 I]]τ2 ,T0 ]] , which follows from the definition of
M (τ ) and the fact that for any t ≥ τ2

E −ã0 I]]τ1 ,τ2 ]] · P̄ τ2 × E E −ã0 I]]τ2 ,T0 ]] · P̄ T0 Ft

(τ1 )
Mt =
E E −ã0 I]]τ1 ,T0 ]] · P̄ T Fτ1
0
E −ã0 I]]τ1 ,τ2 ]] · P̄ τ × E E −ã0 I]]τ2 ,T0 ]] · P̄ T Fτ2

(τ ) .
= 2
0
Mt 2
E E −ã0 I]]τ1 ,T0 ]] · P̄ T0 Fτ1
(τ ) (τ2 )
= Mτ2 1 Mt ,
(τ )
Note here that Mτ2 2 = 1.
We have the following corollaries
Corollary 2.1. Let Assumption 2.1 to be satisfied, then Ψ is a bounded

RCLL semimartingale satisfying
1 ≤ Ψ(t) ≤ C < ∞
for some positive constant C.
Corollary 2.2. Let Assumption 2.1 be satisfied and let (z1∗ , z2∗ ) be given as
in Lemma 2.3, then for any τ ∈ TT0 and for any t ≥ τ
0
E τ(z1∗ , z2∗ )t Ψ(t) = Ψ(τ )E −ã0 I]]τ,T0 ]] · P̄ t , Q0 -a.s.

Zτ,t
Definition. Let Assumption 2.1 be satisfied and let (z1∗ , z2∗ ) be as in

Lemma 2.3. Let
Z t Z tZ
Mt∗ = z1∗ (u)0 dW̃ (u) + z2∗ (u, y)(µ(du, dy) − ν̃(du, dy)),
0 0 E
then M ∗ is a Q0 -local martingale, which is called the variance-optimal

martingale. We introduce an increasing sequence of stopping times
(τn∗ )n∈N defined by
∗
= inf t > τn∗ ∆Mt∗ = −1 ∧ T0 .

τn+1
Remark. The following properties are straightforward

τ
(1) E τ(z1∗ , z2∗ )t = E (M ∗ − (M ∗ ) )t ;
∗ ∗ τ ∗
(2) E τn(z1∗ , z2∗ ) = E τn (z1∗ , z2∗ ) n+1 ;
∗
(3) E τn(z1∗ , z2∗ )t 6= 0 for t ∈ [[τn∗ , τn+1
∗ ∗
[[ on {τn+1 < T }.
We easily derive that

Z t
Zτ0n∗ ,t =1+ Zτ0n∗ ,u− φ(u)0 dW (u)
∗
τn
Z tZ
Zτ0n∗ ,u− ϕ(u, y) − 1 µ(du, dy) − ν(du, dy) .

+
∗
τn E
∗
For simplicity, we introduce Zτ∗n∗ ,t := E τn(z1∗ , z2∗ )t and Z̃τn∗ ,t :=
∗ 0

Ψ(τn )E −ã I]]τ,T0 ]] · P̄ t , so that
Z t Z t Z
Zτ∗∗ ,t = 1 + Zτ∗∗ ,u− z1∗ (u)0 dW̃ (u) + Zτ∗∗ ,u− z2∗ (u, y)(µ(du, dy) − ν̃(du, dy))
n ∗ n ∗ n
τn τn E
Z t
Z̃τn∗ ,t = Ψ(τn∗ ) − Z̃τn∗ ,u− â(u)0 S(u)dW̃ (u)
∗
τn
Z t Z
− Z̃τn∗ ,u− â(u)0 Υ(u, y)(µ(du, dy) − ν̃(du, dy)),
∗
τn E
where â = (â1 , · · · , ân )0 and âi (u) = p̄(u−, Ti )ãi (u) for each i = 1, · · · , n.
A lengthy, though not very difficult computation shows that the follow-
ing representations hold: Let
1 − â(u)0 Υ(u, y)
%(u, y) := .
z2∗ (u, y) + 1
As Ψ is strictly positive, also %(u, y) is strictly positive, thus
1 − â(u)0 Υ(u, y)
z2∗ (u, y) = −1
%(u, y)
and
Z
z1∗ (u) =− S(u)−1 Υ(u, y)z2∗(u, y)ϕ(u, y)λ(u, dy)
ZE
1 − â(u)0 Υ(u, y)

−1
=− S(u) Υ(u, y) − 1 ϕ(u, y)λ(u, dy)
E %(u, y)
:= −Iu (â, %),
where S(u)−1 is the inverse of S(u). Therefore,

Z t
0
Ψ(t) = Ψ(τn∗ ) − Ψ(u−) φ(u) − Iu (â, %) + S(u)0 â(u) dW (u)

∗
τn
Z t
2 0
+ Ψ(u−) kφ(u) − Iu (â, %)k + 2â(u) S(u)φ(u) du
∗
τn
Z tZ 2
1 − â(u)0 Υ(u, y)

+ Ψ(u−) ϕ(u, y) (5)
∗
τn E %(u, y)
+2â(u)0 Υ(u, y)ϕ(u, y) − 1 λ(u, dy)du
Z tZ
%(u, y)
+ Ψ(u−) − 1 µ(du, dy)
∗
τn E ϕ(u, y)
for each n. Thus we consider the following backward semimartingale

equation
 Z t
0
Yu− φ(u) − Iu (a, ρ) + S(u)0 a(u) dW (u)

Y = Y



 t 0 −



 Z t 0
2 0
+ Y (a, ρ)k + 2a(u) S(u)φ(u) du



 u− kφ(u) − Iu

 0
Z tZ 2
1 − a(u)0 Υ(u, y)



+ Yu− ϕ(u, y)



0 E ρ(u, y) (6)
0
+2a(u) Υ(u, y)ϕ(u, y) 1 λ(u, dy)du



 −

 Z tZ
ρ(u, y)




 + Yu− − 1 µ(du, dy),
ϕ(u, y)

0 E







YT0 = 1.

A bounded solution of the BSE (5) is a triple (Y, a, ρ) satisfying (5) such
that
(1) a is an Rn -valued predictable process such that for any stopping time
τ ∈ TT0 ,
n Z 2
ai (u)I]]τ,T0 ]] (u)
X
E E − dp̄(u, Ti ) < ∞;
i=1
p̄(u−, Ti ) T0
(2) ρ(u, y) is a strictly positive P-measurable

e function such that for any
τ ∈ TT0
(z1∗ , z2∗ ) ∈ Zexp

2,τ
1 − a(u)0 Υ(u, y)
where z1∗ (u) := −Iu (a, ρ) and z2∗ (u, y) := − 1;
ρ(u, y)
(3) Y is a bounded RCLL semimartingale with 0 < c1 ≤ Y ≤ c2 < ∞ for
some constants c1 and c2 .
Remark. Although ρ(u, y) is strictly positive, z2∗ may be less than −1,
thus E τ(z1∗ , z2∗ ) may be negative. From (5), we have the following lemma:
Lemma 2.4. Under Assumptions 2.1 and (H), (Ψ, â, %) is a bounded so-
lution of the BSE (6).
With these results we get the form of the variance-optimal martingale as a

first step to the solution of the MVH-problem:
Theorem 2.1. Assume Assumptions (H) and 2.1 be satisfied and (Y, a, ρ)
be a bounded solution of the BSE (5), then for any τ ∈ TT0 , (z1∗ , z2∗ ) is the
solution to the minimization problem (3), i.e.,
Z t Z tZ
Mt∗ := z1∗ (u)0 dW̃ (u) + z2∗ (u, y)(µ(du, dy) − ν̃(du, dy))
0 0 E
is the variance-optimal martingale.
2,τ 0 ∗ 0
E τ(z1 , z2 )t ;

Proof. It is shown that for any (z1 , z2 ) ∈ Zexp , Yt Zτ,t Zτ,t Zτ,t
t ∈ [τ,
T0 ] 0 is ∗a local martingale. As Y is bounded, one can easily derive
0
that Yt Zτ,t Zτ,t Zτ,t E τ(z1 , z2 )t ; t ∈ [τ, T0 ] is uniformly integrable, thus
0 ∗ 0
E τ(z1 , z2 )T Fτ = E YT Zτ,T 0 ∗ 0
E τ(z1 , z2 )T Fτ = Yτ .

E Zτ,T Zτ,T Zτ,T Zτ,T Zτ,T
Similar to the proof of Theorem 3.4 of Kohlmann-Xiong-Ye(2007)31,35, one
derives that (z1∗ , z2∗ ) is the solution of the minimization problem (3) for any
τ ∈ TT0 and Yτ is the optimal cost, i.e.,
2
0 2
Yτ = E Zτ,T0
Eτ z1∗ , z2∗ T Fτ = ess inf E 0 τ
Zτ,T0 E (z1 , z2 )T0 Fτ ,
0 2,τ
(z1 ,z2 )∈Zexp
which completes the proof.
Corollary 2.3. Under Assumptions (H) and 2.1, there exists a unique
solution of the BSE (6).
2.2. E ∗ -Martingale and Mean-Variance Hedging

In this section, we assume that
Assumption (BS): the BSE (6) has a bounded solution denoted by

(Y, a, ρ).
Recall that
z1∗ (u) = −Iu (a, ρ),
1 − a(u)0 Υ(u, y)
z2∗ (u, y) = − 1,
ρ(u, y)
where
1 − a(u)0 Υ(u, y)
Z
−1
Iu (a, ρ) = S(u) Υ(u, y) − 1 ϕ(u, y)λ(u, dy).
E ρ(u, y)
From Theorem 2.1
Z t Z tZ
Mt∗ := z1∗ (u)0 dW̃ (u) + z2∗ (u, y)(µ(du, dy) − ν̃(du, dy))
0 0 E
is the variance-optimal martingale. As above, we let

∗
= inf t > τn∗ ∆Mt∗ = −1 ∧ T0 .

τn+1
n ∗
o
Also Zτ0n∗ ,t E(M ∗ − (M ∗ )τn )t ; t ∈ [τn∗ , T0 ] is a square-integrable mar-
tingale, and one derives from Choulli-Krawczyk-Stricker(1998)13 that
Z 0 E(M ∗ ) is a regular stochastic exponential satisfying the reverse Hölder
inequality (R2 ), i.e., there exists a constant C such that
h 2 i
0 ∗ ∗ τ
E Zτ,T 0
E (M − (M ) )T0
Fτ ≤ C < ∞, a.s.
Similar to Definition 3.11 of 13, we give the definition of E ∗ -martingale
Definition.
(i) X is an E ∗ -martingale, if for any n

∗
0 ∗ ∗ τn

E Zτn∗ ,T0 E M − (M )
Xτn < ∞
∗
T0
n ∗
o
and (Xt − Xt∧τn∗ )Zτ0n∗ ,t E M ∗ − (M ∗ )τn t ; t ∈ [0, T0 ] is a martin-
gale. The class of E ∗ -martingales is denoted by M(E ).
∗
(ii) X is called as an E -local martingale if (Xt − Xt∧τn∗ )Zτ0n∗ ,t

∗
∗
E M ∗ − (M ∗ )τn t ; t ∈ [0, T0 ] is a local martingale for each n. The

class of E ∗ -local martingales is denoted by Mloc (E ∗ ).

Lemma 2.5. Under Assumption (BS), an RCLL process X is a E ∗ -local
martingale iff it is a semimartingale such that X + [M ∗ , X] is a Q0 -local
martingale.
∗
∗ ∗ ∗ τn

Proof. X is a0 local E -martingale iff (Xt −Xt∧τn )E M − (M ) t ; t ∈
∗
[0, T0 ] is a Q -local martingale for each n. The rest is completely the same
as in the proof of Proposition 3.15 of 13.
Proposition 2.1. Let Assumption (BS) be satisfied and let X be a special

semimartingale, then X is an E ∗ -local martingale iff there exist an Rd -
valued predictable process ξ1 and a P-measurable
e function ξ2 (u, y) such that
Z t Z tZ
Xt = X0 + ξ1 (u)0 dW (u) + ξ2 (u, y)µ(du, dy)
Z t0 0 E
− ξ1 (u)0 {z1∗ (u) + φ(u)}du

0
Z tZ
ξ2 (u, y) z2∗ (u, y) + 1 ϕ(u, y)λ(u, dy)du,

−
0 E
0
Q -a.s., for every t ∈ [0, T0 ].
Proof. Prove that Xt + [X, M ∗ ]t is a Q0 -local martingale iff

t t
Z Z Z
ξ1 (u)0 {z1∗ (u) + φ(u)}du + z2∗ (u, y) + 1 ϕ(u, y) − 1 λ(u, dy)du = 0,

At + ξ2 (u, y)
0 0 E
Q0 -a.s., for each t ∈ [0, T0 ], and hence

Z t Z t Z
Xt = X0 + ξ1 (u)0 dW (u) + ξ2 (u, y)µ(du, dy)
Z 0t 0 E Z t Z

− ξ1 (u)0 {z1∗ (u) + φ(u)}du − ξ2 (u, y) z2∗ (u, y) + 1 ϕ(u, y)λ(u, dy)du,
0 0 E
which completes the proof.
Corollary 2.4. Let Assumption (BS) be satisfied and let H be any contin-
gent claim belonging to L2 (FT0 , P ), then the BSE
 Z t Z tZ
0


 h t = h 0 + ξ1 (u) dW (u) + ξ2 (u, y)µ(du, dy)
 0 0 E
Z t Z t Z

 − ξ1 (u)0 {z1∗ (u) + φ(u)}du − ξ2 (u, y) z2∗ (u, y) + 1 ϕ(u, y)λ(u, dy)du,

 0 0 E

hT = H
(7)
has a solution (h, ξ1 , ξ2 ).
Proof. Let h be an RCLL process on [0, T0 ] satisfying

∗
EQ0 HE(M ∗ − (M ∗ )τn )T0 Ft

ht = ∗
E(M ∗ − (M ∗ )τn )t
for τn∗ ≤ t < τn+1
∗
and for each n. As Z 0 E(M ∗ ) satisfies the reverse Hölder
inequality (R2 ) and H ∈ L2 (FT0 , P ), one directly derives from the proof
of Proposition 3.12(iii) of Choulli-Krawczyk-Stricker(1998)13 that h is an
E ∗ -martingale. From Theorem 4.1 of 13, we see that

max |ht | ≤ CkHk2 < ∞
2 t∈[0,T0 ]
for some constant C, thus h is a special semimartingale. It follows from

Proposition 2.1 that there exist an Rd -valued predictable process ξ1 and a
P-measurable
e function ξ2 (u, y) such that (h, ξ1 , ξ2 ) is a solution of the BSE
(7).
With these results mainly adopted from recent results on mean variance
hedging we now turn to the already above stated mean-variance hedging
problem in the bond market and derive an analytical solution. Given a
contingent claim H ∈ L2 (FT0 , P ), which might not be fully replicated by
a self-financing portfolio based on the maturities T1 , · · · , Tn , we will try to

solve the following minimization question
h 2 i
Problem (MVH-n): J= min E H − VTx,π 0
.
π∈A(T1 ,··· ,Tn )
To give an explicit form of the optimal strategy, we need to introduce

ϕ(u, y)
Z
<u := S(u)S(u)0 + Υ(u, y)Υ(u, y)0 λ(u, dy) and
ZE ρ(u, y)
ϕ(u, y)
ζu := S(u)ξ1 (u) + Υ(u, y)ξ2 (u, y) λ(u, dy).
E ρ(u, y)
Since S(u) is a n×n invertible matrix for each u and both ϕ(u, y) and ρ(u, y)
are strictly positive, <u is a symmetric matrix which is strictly positive. We
have the following theorem
Theorem 2.2. Let Assumptions (H) and 2.1 be satisfied and let (Y, a, ρ)
be the bounded solution of the BSE (6). For a given contingent claim H ∈
L2 (FT0 , P ), let (h, ξ1 , ξ2 ) be a solution of the BSE (7), let
∗
x,π
π ∗ (u) := <−1
u ζu + (hu− − Vu− )a(u), (8)
then π ∗ is the optimal strategy of Problem (MHV-n). The optimal cost is
given by
J = Y0−1 (h0 − x)2
Z T0
ϕ(u, y)
Z
−1
+E Yu− kξ1 (u)k2 + ξ2 (u, y)2 λ(u, dy) − ζu0 <−1
u u du .
ζ
0 E ρ(u, y)
∗
Remark. We let Vt∗ := Vtx,π be the wealth process corresponding to
(x, π ∗ ), one can see that h − V ∗ is the solution of the following stochastic
differential equation (SDE)
Z t
∗
ht − Vt∗ = h0 − x + Rt − (hu− − Vu− )dM̃u , (9)
0
where
Z t
ξ1 (u)0 − ζu0 <−1 ∗

Rt := u S(u) dW (u)
0 Z Z
t
+ {ξ2 (u, y) − ζu0 <−1 ∗
u Υ(u, y)}(µ(du, dy) − ν (du, dy)),
Z t0 E Z t
M̃t := − a(u)0 S(u)dW ∗ (u) − a(u)0 Υ(u, y)(µ(du, dy) − ν ∗ (du, dy))
0 0
and where
Z t
W ∗ (t) = W (t) − {z1∗ (u) + φ(u)}du
0
ν ∗ (du, dy) = z2∗ (u, y) + 1 ϕ(u, y)λ(u, dy)du.

To solve the SDE (9), we let Z̃t := E(M̃ )t , one can see that Z̃τn∗ = 0 and
26
Z̃− I[[τn∗,τn+1
∗
[[ 6= 0 for each n. It follows from Theorem 6.8 of Jacod(1979)
that the solution of the SDE (9) can be written as
X
h−V∗ = (h − V ∗ )(n) I[[τn∗ ,τn+1
∗ [[ ,
n≥0
where

n 1 ∗ ∗
(h − V ∗ )(n) = Z̃ ∆Rτn∗ + n · (Rτn+1 − Rτn )
Z̃−
1 h ∗ ∗
i
τn+1 τn
+ n I[[0,τn+1 ∗ [[ · R, M̃ − M̃
Z̃−
n ∗ ∗ ∗
and where Z̃ = E(M̃ − M̃ τn ) = E(M̃ τn+1 − M̃ τn ).
Proof of Theorem 2.2. First prove that
x,π 2
Yt−1 (ht − Vt ) = Y0−1 (h0 − x)2 + mt
Z t 2
−1 0 x,π 0
+ Yu− ξ1 (u) − S(u) π(u) + (hu− − Vu− )S(u) a(u) du
Z0t Z n
−1 0 o2 ϕ(u, y) x,π
+ Yu− ξ2 (u, y) − π(u) Υ(u, y) + hu− − Vu− λ(u, dy)du
Z0t ZE ρ(u, y)
−1 x,π 0 0 ϕ(u, y)
+ 2Yu− hu− − Vu− π(u) Υ(u, y) − ξ2 (u, y) {1 − a(u) Υ(u, y)} λ(u, dy)du
Z0t ZE ρ(u, y)
n 2 o ϕ(u, y)
−1 x,π 2 0
− Yu− hu− − Vu− 1 − a(u) Υ(u, y) λ(u, dy)du,
0 E ρ(u, y)
where
Z t
x,π 0
mπ
t =
−1
Yu− 2 hu− − Vu− ξ1 (u) − S(u)0 π(u)
0
x,π 2 0
+ hu− − Vu− φ(u) − Iu (a, ρ) + S(u)0 a(u) dW (u)
Z t Z

−1
+ Yu− ξ2 (u, y) − π(u)0 Υ(u, y)
0 E
x,π 2 ϕ(u, y) x,π 2
+ hu− − Vu− − hu− − Vu− {µ(du, dy) − ν(du, dy)}
ρ(u, y)
is a local martingale under P . Therefore,

x,π
Yt−1 (ht − Vt )2 = Y0−1 (h0 − x)2 + mπ t
Z t 2
−1 0 x,π 0
+ Yu− ξ1 (u) − S(u) π(u) + (hu− − Vu− )S(u) a(u) du
Z0t Z 2
−1 0 x,π 0 ϕ(u, y)
+ Yu− ξ2 (u, y) − Υ(u, y) π(u) + hu− − Vu− Υ(u, y) a(u) λ(u, dy)du
0 E ρ(u, y)
= Y0−1 (h0 − x)2 + mπ

Z t t 0
−1 −1 x,π −1 x,π
+ Yu− π(u) − <u ζu − (hu− − Vu− )a(u) <u π(u) − <u ζu − (hu− − Vu− )a(u)
0 Z
2 ϕ(u, y) 0 −1
+kξ1 (u)k2 + ξ2 (u, y) λ(u, dy) − ζu <u ζu du.
E ρ(u, y)
As π ∈ A(T1 , · · · , Tn ), we derive that V x,π is an E ∗ -local martingale with

VTx,π
0
∈ L2 (P ), thus V x,π is in fact an E ∗ -martingale. It follows from Theo-
rem 4.1 (ii) of Choulli-Krawczyk-Stricker(1998)13 that
max |Vtx,π | ≤ kV x,π k2 < ∞.

2 T0
t∈[0,T0 ]
Similarly, one can see that

max |ht | ≤ kHk2 < ∞.
t∈[0,T0 ] 2
Since Y is a bounded process with 0 < c1 ≤ Y ≤ c2 < ∞ for some constants

c1 and c2 , one can see that Y −1 (h − V x,π )2 is a submartingale which is
belongs to the class (D), thus m is a uniformly integrable martingale and
Z t
0
−1 x,π x,π
Yu− π(u) − <−1
u ζu − (hu− − Vu− )a(u) <u π(u) − <−1
u ζu − (hu− − Vu− )a(u)
0 Z
ϕ(u, y)
+kξ1 (u)k2 + ξ2 (u, y)2 0
λ(u, dy) − ζu <−1
u ζu du
E ρ(u, y)
is an integrable increasing process. Therefore,
x,π 2 −1 x,π 2
E(hT − V ) = EY (hT − V )
0 T0 T0 0 T0
−1
=Y (h0 − x)2
0
Z T 0
0 −1 −1 x,π −1 x,π
+E Y π(u) − <u ζu − (hu− − V )a(u) <u π(u) − <u ζu − (hu− − V )a(u)
u− u− u−
0 Z
2 ϕ(u, y) 0 −1
+kξ1 (u)k2 + ξ2 (u, y) λ(u, dy) − ζu <u ζu du.
E ρ(u, y)
From the definition of π ∗ in (8), one can see that

∗
Yt−1 (ht − Vtx,π )2 = Y0−1 (h0 − x)2 + mπt
∗
Z t
2 ϕ(u, y)
Z
−1 2 0 −1
+ Yu− kξ1 (u)k + ξ2 (u, y) λ(u, dy) − ζu <u ζu du.
0 E ρ(u, y)
∗
Since mπ is only a local martingale, there exists a sequence of stopping
∗
times (τn ) with ↑ limn→∞ τn = T0 such that {mπt∧τn ; t ∈ [0, T0 ]} is a uni-
∗ 2
formly integrable martingale for each n. Since Y h − V x,π is left con-
tinuous at T0 , one can see from Fatou’s lemma that
2 2 2
x,π ∗ x,π ∗ x,π ∗
E hT0 − VT = E YT0 hT0 − VT =E lim inf Yτn hτn − Vτ
0 0 n
n→∞
x,π ∗ 2
≤ lim inf E Yτn hτn − Vτn
n→∞
Z τn
Z
−1 2 2 ϕ(u, y) 0 −1
= Y0−1 (h0 − x)2 + lim inf E Yu− kξ1 (u)k + ξ2 (u, y) λ(u, dy) − ζu <u ζu du
n→∞ ρ(u, y)
Z T 0 Z E
0 −1 2 2 ϕ(u, y) 0 −1
≤ Y0−1 (h0 − x)2 + E Yu− kξ1 (u)k + ξ2 (u, y) λ(u, dy) − ζu <u ζu du
2 0 E ρ(u, y)
x,π
≤E hT0 − VT ,
0
from which we can see that π ∗ ∈ A(T1 , · · · , Tn ) and π ∗ is the optimal

strategy.
2.3. The Case of not Invertible S

In this section, we consider the mean-variance hedging when σ(u, T ) is
independent of the maturity time T , i.e., σ(u, T ) ≡ σ(u), where {σ(u); u ∈
[0, T ∗ ]} is a bounded Rn -valued predictable process, and δ(u, y, T ) is of the
following form
δ̃(u, y)
δ(u, y, T ) = −
1 + δ̃(u, y)(T − u)
1
e function δ̃(u, y) with δ̃(u, y) > − ∗ . One can
T
see that S(u, T ) = −σ(u)(T − u) and eD(u,y,T ) − 1 = δ̃(u, y)(T − u), thus
the discounted bond price maturing at T is given by
Z ·
p̄(t, T ) = p̄(0, T )E − σ(u)0 (T − u)dW̃ (u)
0
Z ·Z
+ δ̃(u, y)(T − u)(µ(du, dy) − ν̃(du, dy)) .
0 E t
(10)
Fix a T0 ∈ (0, T ∗ ) and given any integer n0 , let T1 , · · · , Tn0 be n0
maturities with T0 < Ti < T ∗ for each i and T1 < T2 < · · · < Tn0 . Given
a contingent claim H ∈ L2 (FT0 ), we consider the mean-variance hedging
problem of H by a self-financing portfolio based on p̄(t, T1 ), . . . , p̄(t, Tn0 ).
From the following Lemma, one can see that to hedge H, we only need to
trade the bond maturing at T1 :
Lemma 2.6. Assume that the discounted bond price p̄(t, T ) is given by
(10), then for any
π = (π1 , · · · , πn0 )0 ∈ A(T1 , · · · , Tn0 )
there exists π̂ ∈ A(T1 ) such that for any t ∈ [0, T0 ]
Vtx,π = Vtx,π̂ , a.s.,
thus for a given contingent claim H ∈ L2 (FT0 , P ),

h 2 i h 2 i
min E H − VTx,π0
= min E H − VTx,π
0
.
π∈A(T1 ,··· ,Tn0 ) π∈A(T1 )
Proof. One can see from (1) that

n0
Z tX
Vtx,π = x − πj (u)(Tj − u)σ(u)0 dW̃ (u)
0 j=1
Z tZ X n0

+ πj (u)(Tj − u)δ̃(u, y) µ(du, dy) − ν̃(du, dy)
0 E j=1
Z t
=x− π̂(u)(T1 − u)σ(u)0 dW̃ (u)
0
Z tZ

+ π̂(u)(T1 − u)δ̃(u, y) µ(du, dy) − ν̃(du, dy)
0 E
= Vtx,π̂
n0
X Tj − u
where π̂(u) = πj (u) belongs to A(T1 ).
j=1
T1 − u
Thus it is natural to consider the following problem

h 2 i
Problem (MHV-1): J = minπ∈A(T1 ) E H − VTx,π 0
.
2,τ
We now denote Zd exp for the family of (z1 , z2 ) satisfying (6) with
0 τ
2
E Zτ,T0 E (z1 , z2 )T0 < ∞, and consider the following kind of mini-
mization problem
h
0
2 i
Ψ(τ ) := ess inf E Zτ,T 0
E τ(z1 , z2 )T0 Fτ . (11)
d2,τ
(z1 ,z2 )∈Zexp
Note that the adjustment process ã in this case is only a 1-dimensional

process. Instead of Assumption 2.1, we here directly assume that
Assumption 2.2. There exists a pair (z1∗ , z2∗ ) satisfying (2) such that for
2,τ
all τ ∈ T , (z ∗ , z ∗ ) ∈ Zd
T0 1 and
2 exp
h
0 τ ∗ ∗
2 i
Ψ(τ ) = EQ0 Zτ,T0
E (z , z )
1 2 T0 Fτ , Q0 -a.s.,
and
Z t Z tZ
Mt∗ = z1∗ (u)0 dW̃ (u) + z2∗ (u, y)(µ(du, dy) − ν̃(du, dy))
0 0 E
is again called the variance-optimal martingale.
To describe the variance-optimal martingale, we consider the following

BSE which is similar to (6)
 Z t


 Y t = Y 0 − Yu− {φ(u) + ẑ1 (u) − a(u)σ(u)}0 dW (u)
 0
t
 Z Z

 ϕ(u, y) 2

 + Y u− 1 − a(u)δ̃(u, y) + 2a(u)δ̃(u, y)ϕ(u, y) − 1 λ(u, dy)du

 Z0 E ρ(u, y)
t
 + Yu− kφ(u) + ẑ1 (u)k2 − 2a(u)σ(u)0 φ(u) du

 Z0 t Z

 ρ(u, y)

 + Yu− − 1 µ(du, dy),


 0 E ϕ(u, y)

YT0 = 1.
(12)
A bounded solution of the BSE (12) is a tuple (Y, a, ẑ1 , ρ) satisfying (12)
such that
(1) a is an Rn -valued predictable process such that for all stopping times
τ ∈ TT0 ,
2
a(u)I]]τ,T0 ]] (u)
Z
E E − dp̄(u, T1 ) < ∞;
p̄(u−, T1 )(T1 − u) T0
(2) ẑ1 is an Rn -valued W -integrable predictable process and ρ(u, y) is a

strictly positive P-measurable
e function such that for any τ ∈ TT0
2,τ
(ẑ1 , ẑ2 ) ∈ Zexp
d
1 − a(u)δ̃(u, y)
where ẑ2 (u, y) := − 1;
ρ(u, y)
(3) Y is a bounded RCLL semimartingale with 0 < c1 ≤ Y ≤ c2 < ∞ for
some constants c1 and c2 .
Similar to Lemma 2.4 and Theorem 2.1, we can show the following
theorem
Theorem 2.3. 1) Under Assumption 2.2, the BSE (12) has a bounded
solution;
2) If the BSE (12) has a bounded solution denoted by (Y, a, ẑ1 , ρ), then
Yt = Ψ(t), a.s., for all t ∈ [0, T0 ] and
Z t Z tZ
Mt∗ := ẑ1 (u)0 dW̃ (u) + ẑ2 (u, y)(µ(du, dy) − ν̃(du, dy))
0 0 E
is the variance-optimal martingale.
We assume that the BSE (12) has a bounded solution denoted by

(Y, a, ẑ1 , ρ), let
Z t Z tZ
Mt∗ := ẑ1 (u)0 dW̃ (u) + ẑ2 (u, y)(µ(du, dy) − ν̃(du, dy))
0 0 E
be the variance-optimal martingale. In a similar way as above we can define
E ∗ -martingales and E ∗ -local martingales in this case and show that the
following BSE always has a solution for any given H ∈ L2 (FT0 , P ):
 Z t Z tZ
 ht = h0 +

 ξ1 (u)0 dW (u) + ξ2 (u, y)µ(du, dy)
 Z 0 0 E Z Z
t t
 − ξ1 (u)0 {ẑ1 (u) + φ(u)}du − ξ2 (u, y) ẑ2 (u, y) + 1 ϕ(u, y)λ(u, dy)du,

 0 0 E

hT = H
(13)
has a solution (h, ξ1 , ξ2 ).
Let
ϕ(u, y)
Z
ˆ u = kσ(u)k2 +
< δ̃(u, y)2 λ(u, dy) and
E ρ(u,
Z y)
ϕ(u, y)
ζ̂u = −σ(u)0 ξ1 (u) + δ̃(u, y)ξ2 (u, y)λ(u, dy).
E ρ(u, y)
In a similar way as in Theorem 2.2, we can derive the following theorem
Theorem 2.4. Given a bounded Rn -valued predictable process {σ(u); u ∈

1
[0, T ∗ ]} and a bounded P-measurable
e function δ̃(u, y) with δ̃(u, y) > − ∗
T
satisfying < ˆ u ≥ c > 0 for some positive constant c, assume that the dis-
counted bond price maturing at T is given by (10). Let Assumptions 12 be
satisfied and let (Y, a, ρ) be the bounded solution of the BSE (12). For a
given contingent claim H ∈ L2 (FT0 , P ), let (h, ξ1 , ξ2 ) be a solution of the
BSE (13), let

∗ 1 ζ̂u x,π ∗
π̂ (u) := + (hu− − Vu− )a(u) , (14)
T1 − u <û
then π ∗ is the optimal strategy for Problem (MHV-1). The optimal cost is
then given by
Z T0
J = Y0−1 (h0 − x)2 + E −1
Yu− kξ1 (u)k2
0
ζ̂u2

2 ϕ(u, y)
Z
+ ξ2 (u, y) λ(u, dy) − du .
E ρ(u, y) û
<
Remark. The reader should note that in Lemma (2.6) T1 can be replaced
by any Ti , i ∈ {1, · · · , n0 }. So the agent is free to choose the bond with any
admissible time of maturity. As a consequence, it is not surprising that the
optimal cost is independent of the chosen maturity time what is directly
seen from Theorem 2.4. A similar observation in a simpler market setting is
already made in Proposition 4.5 in Björk (1997)5 . We should however note
that this is somewhat contradictory to observations in reality: There seems
to be an extra risk in hedging a claim with exercise time one month with a
bond expiring in ten years (see (17).
3. Optimal Exponential Utility in Jump Bond Markets

Still we consider the continuous-time bond market with jumps driven by an
n-dimensional Brownian motion W = (W1 , · · · , Wn )0 as well as an integer-
valued random measure µ(du, dy) on R+ × R with compensator ν(du, dy) =
λu (dy)du as in section 1. However we need some few more assumptions.
The main difficulty arises from the fact stated in 7 that ”in the continuous-
time bond market model there is naturally a continuum of basic traded
securities (zero-coupon bonds parameterized by their maturities) while in
the standard model of stock market there is normally only a finite number
of securities.” So, again, we first construct an arbitrage-free market with
given parameters (σ, δ, φ, ϕ) such that
(i) σ(u, T ) is a bounded Rn -valued predictable process on [0, T ], and

δ(u, y, T ) is a bounded P-measurable
e function on [0, T ] × E for ev-
∗
ery fixed T ∈ [0, T ]; We assume that both of σ(u, T ) and δ(u, y, T )
areR continuously differential in the T -variable, let D(u, y, T ) :=
T
− u δ(u, y, l)dl;
(ii) φ is a bounded Rn -valued predictable process on [0, T ∗];
(iii) ϕ is a strictly positive P-measurable
e function on [0, T ∗] × E with 0 <
k1 ≤ ϕ(u, y) ≤ k2 < ∞ for two constants k1 and k2 .
Again, let
Z t
f (t, T ) = f (0, T ) + Ψ(u, T ; σ, δ, φ, ϕ)du
Z t 0 Z tZ (15)
0
+ σ(u, T ) dW (u) + δ(u, y, T )µ(du, dy),
0 0 E
where
Z T
Ψ(u, T ; σ, δ, φ, ϕ) := σ(u, T ) 0
σ(u, s)ds − σ(u, T )0 φ(u)
u Z
− δ(u, y, T )eD(u,y,T )ϕ(u, y)λu (dy).
E
(16)
dQ0

The bond market is arbitrage-free, since Q0 defined by dP F := ZT0 ∗ is a
equivalent martingale measure, where
Z t Z tZ
0
Zt0 = 1 + φ(u)0 dW (u) + 0
Zu− Zu− ϕ(u, y) − 1 µ(du, dy) − ν(du, dy) ,
0 0 E
is a strictly positive uniformly integrable martingale on [0, T ∗ ], i.e.,

for R each maturity T , R the discounted
RT
bond price process p̄(t, T ) =
t t
e− 0 ru du p(t, T ) (:= e− 0 ru du e− t f (t,s)ds ) is a local martingale on [0, T ]
under Q0 . Of course, there may be many other equivalent martingale mea-
sure. In this paper, we consider the so called ’minimal entropy martingale’
(the density process of the minimal entropy martingale measure) based on
the bonds maturing in [T0 , T ∗ ] and then consider the following exp-utility
U (x) = − exp{−k0 x} optimization problem:
Problem (exp-utility opt): sup EU (XT0 )
X∈W(x)
where W(x) is the set of all admissible self-financing approximate wealth

processes based on the bonds maturing in [T0 , T ∗ ] with initial wealth x.
Then we consider the dynamic exp utility indifference value process for a
given H ∈ L∞ (FT0 ), which is defined as a bounded RCLL semimartingale
C(H) = {Ct (H); t ∈ [0, T0 ]} satisfying

ess sup E U (XT0 − Xt )Ft = ess sup E U Ct (H) + XT0 − Xt − H Ft
X∈W X∈W
38
(see Mania and Schweizer(2005) ). Similar problems have been considered
in Becher(2006)3, Mania and Schweizer(2005)38, or Morlais(2008)40. The
main difference between this paper and those papers is that we have a
continuum of traded securities and the wealth process is described by an
(M, Q0 )-normalized martingale.
The minimal entropy martingale measure (MEMM) has been investi-

gated in various settings by several authors among them: Delbaen et al.
(2002)22 , Frittelli (2000)23 , Fujiwara and Miyahara (2003)24 , Kohlmann
and Xiong (2008)33 , Mania, Santacroce and Tevzadze (2003)37 , Mania and
Schweizer (2005)38 , etc. But in the bond market, as the system of bonds
maturing in [T0 , T ∗ ] is a continuum of securities, it is tempting to consider
portfolios given by measures on maturities as in [6] or [7]. A crucial diffi-
culty then arises from the fact that this set of portfolios is not complete. So
any limiting procedure will make us leave a given set of admissible strate-
gies and corresponding wealths, that is, by taking limits we always arrive
at only approximate values. So, in a natural setting the notions of approx-
imate completeness, approximately attainable claim, approximate wealth
etc are found.
De Donno and Pratelli(2004)19 consider the theory of cylindrical inte-
gration and the normalized integral, which was first introduced in Mikule-
vicius and Rozovskii(1998)39, to generalize the concept of measure-valued
strategies as measure valued portfolios are not sufficient to describe all fi-
nancial portfolios (see the discussion in De Donno and Pratelli(2004)19 and
De Donno and Pratelli(2005)20). In this paper, similar to De Donno and
Pratelli(2004)19, we introduce the concept of ’(M, Q0 )-normalized martin-
gale’ and view the local (M, Q0 )-normalized martingale as an approximate
wealth process. As a main contribution to the mathematical theory we add
to the technical results in De Donno and Pratelli(2004)19 a representa-
tion for a given (M, Q0 )-normalized martingale and describe the dynam-
ics of the minimal entropy martingale (MEM) based on the bonds matur-
ing in [T0 , T ∗ ] by a backward semimartingale equation (BSE) based on an
(M, Q0 )-normalized martingale, see Section 3. Then we give an example
in Section 4, in which we can give the explicit solution to the BSE, and
in Section 5, we give the explicit form of the optimal approximate wealth
for the Problem (exp-utility opt) by using the solution of the BSE. Similar
problems are considered in De Donno and Pratelli(2005)20 in a different
setting based on generalized strategies. Here we make direct use of the nor-
malized martingales instead which turns out to be advantageous in hedging
problems.
Finally, in Section 6 we discuss the exp utility indifference valuation
(exp UIV) of a bounded contingent claim H ∈ L∞ (FT0 ), we introduce the
concept of ’exp-orthogonality’ and then describe the dynamics of the exp
UIV of H by another BSE under the minimal entropy martingale measure
Q∗ in the incomplete market.
We are aware of the fact that from a practical point of view our results
are not very satisfying. However, in our minds, the results might bring for-
ward the general theory of bond markets which still needs a lot of research.
For a given maturity T0 , let Z be the set of strictly positive martingales
on [0, T0 ] with Z0 = 1 such that for any maturity T ∈ [T0 , T ∗ ], Zt p̄(t, T )
is a local martingale on [0, T0 ]. Then we know that for any Z ∈ Z, there
exists a pair (z1 , z2 ) satisfying (5) for all T ∈ [T0 , T ∗ ] and
z1 ,z2 0 ez1 ,z2
Zt = Z
t Z := Zt Zt
·
=E (φ(u) + z1 (u))0 dW (u)
0 Z ·Z

+ ϕ(u, y) 1 + z2 (u, y) − 1 (µ(du, dy) − ν(du, dy)) .
0 E t
Definition. A martingale Ẑ ∈ Z is called the minimal entropy martingale

(MEM) based on [T0 , T ∗ ] if it is the solution of the following optimization
problem:

min E ZT ln ZT .
Z∈Z
Remark. One can easily see that Z 0 ∈ Z satisfies E ZT0 ln ZT0

< ∞.
Furthermore, since D(u, y, T ) is bounded for each fixed T ∈ (0, T ∗ ], one gets
from (3) that p̄(·, T ) is a locally bounded process for each fixed T . From
Theorem 2.1 and Theorem 2.2 of Frittelli(2000)23 that a MEM based on
[T0 , T ∗ ] always exists! In this paper, we will describe Ẑ by a backward
semimartingale equation based on the normalized martingale.
3.1. The Minimal Entropy Martingale Based on [T0 , T ∗ ]

In this section, we will describe the dynamics of the MEM by a backward
semimartingale equation. As usual, we denote by Ẑ the minimal entropy
martingale based on [T0 , T ∗ ], and as we have seen, there exists a pair (z1∗ , z2∗ )
satisfying (5) for all T ∈ [T0 , T ∗ ] and
z ∗ ,z2∗ ∗ ∗
etz1 ,z2 .
Ẑt = Zt 1 := Zt0 Z
∗ ∗
As we do not want (z1∗ , z ∗ ) be too ’strange’ to be able to write Zez1 ,z2 as a
stochastic exponential, we need the following assumption
Assumption 3.1. Assume that (z1∗ , z ∗ ) satisfy

Z T0 Z T0 Z
kz1∗ (u)k2 du + ϕ(u, y)z2∗ (u, y)2 λu (dy)du < ∞, Q0 -a.s..
0 0 E
In other words,

ez1∗ ,z2∗ = E M z1∗ ,z2∗ ,
Z
Z t Z tZ
z ∗ ,z2∗
where Mt 1 := z1∗ (u)0 dW
f (u) + z2∗ (u, y)(µ(du, dy) − ν̃(du, dy)) is
0 0 E
a locally square integrable martingale under Q0 .
It is natural to introduce
 
z1 ,z2

 Z ∈ Z with Z0 = 1 and E ZT0 ln ZT0 < ∞ and 

ZEnt := Z z1 ,z2 .
Z T Z T Z
0 0
kz1 (u)k2 du + ϕ(u, y)z2 (u, y)2 λu (dy)du < ∞, Q0 -a.s. 


 

0 0 E
We have the following lemma
Lemma 3.1. Z 0 belongs to ZEnt satisfies the reverse Hölder inequality

REnt (P ):
0 0

E Zτ,T 0
ln Zτ,T0
Fτ ≤ k < ∞, a.s.
0
ZT00
for all stopping times τ with τ ≤ T0 , where Zτ,T = and k is a positive
0
Zτ0
constant.
∗ ∗
Remark. Obviously, under Assumption 3.1 the MEM Ẑ = Z z1 ,z2 is also
the solution of the following optimization problem:
E ZTz1 ,z2 ln ZTz1 ,z2 .

Problem (MEM): min
z ,z Z 1 2 ∈Z
Ent
Note that now we can trade all bonds with maturities in [T0 , T ∗ ]!
Since there exists a continuum of basic traded securities, we introduce the
cylindrical martingale and the normalized martingale which is related to
the normalized integral in De Donno and Pratelli(2004)19, which was first
introduced in Mikulevicius-Rozovskii(1998)39 as a mathematical tool with-
out relation to financial problems.
3.2. The Locally Square Integrable (LSI) Cylindrical

Martingales and the Normalized Martingale
n o
M T t∈[0,T0 ] ; T ∈ [T0 , T ∗ ] of square integrable

We consider the family
martingale under Q0 , where M T is given by
Z t Z tZ
MtT := S(u, T )0 dW
D(u,y,T )
f (u) + e − 1 (µ(du, dy) − ν̃(du, dy)),
0 0 E
one can see that for any T1 , T2 ∈ [T0 , T ∗ ],

D E Z Z
t
0 D(u,y,T1 ) D(u,y,T2 )
M T1 , M T2 = S(u, T1 ) S(u, T2 ) + e −1 e − 1 ϕ(u, y)λu (dy) du
t Z0t E
= Qu (T1 , T2 )du,
0
where
Z
0
eD(u,y,T1 ) −1 eD(u,y,T2 ) −1 ϕ(u, y)λu (dy).

Qu (T1 , T2 ) = S(u, T1 ) S(u, T2 )+
E
Let C = C([T0 , T ]), the space of all continuous function on [T0 , T ∗ ],

∗
then its topological dual R is the space of all Radon measures on [T0 , T ∗ ].
Note that (C, R) is a Schwartz pair. We have the following lemma:
Lemma 3.2. For any Radon measure ` ∈ R, let

Z t Z tZ
` 0 f
Mt := `(S(u, ·)) dW (u) + `(eD(u,y,·) − 1)(µ(du, dy) − ν̃(du, dy))
0 0 E
where
T∗ T∗ T∗
Z Z Z 0
`(S(u, ·)) := S(u, T )`(dT ) = S1 (u, T )`(dT ), · · · , Sn (u, T )`(dT ) ,
ZTT0 ∗ n o
T0 T0
`(eD(u,y,·) − 1) := eD(u,y,T ) − 1 `(dT ),

T0
then the family of random processes M = M ` ; ` ∈ R is a locally square

integrable (LSI) cylindrical martingale with covariance operator function Q

(see Mikulevicius and Rozovskii(1998)39 ), i.e., for any `1 , `2 ∈ R,
Z t

`1 `2

M ,M t = h`1 , Qu `2 idu
0
Z
0
`1 eD(u,y,·) − 1 `2 eD(u,y,·) −

where h`1 , Qu `2 i = `1 (S(u, ·)) `2 (S(u, ·)) +
E
1 ϕ(u, y)λu (dy).
Similar to De Donno and Pratelli(2004)19, we consider R-valued pre-

dictable processes: An R-valued predictable process ` = `t (dT ); t ∈
R T∗
[0, T0 ] is a process such that for each f ∈ C, T0 f (T )`t (dT ); t ∈ [0, T0 ]
is a predictable process. We denote by L2 (M, R) the set of R-valued pre-
dictable processes ` satisfying
Z T0
EQ0 hù , Qu ù idu
0 Z T Z
0 2
kù (S(u, ·))k2 + ù eD(u,y,·) − 1

:= EQ0 ϕ(u, y)λu (dy) du < ∞.
0 E
Note that different from De Donno and Pratelli(2004)19, for any ` ∈

L2 (M, R), we can define the stochastic integral directly by
Z t Z t
It (`) = ù dM(u) : = ù (S(u, ·))0 dW
f (u)
0 0 Z tZ
+ ù (eD(u,y,·) − 1)(µ(du, dy) − ν̃(du, dy)),
0 E
Z t
and by stopping we can extend the stochastic integral ù dM(u) to
0
any ` ∈ L2loc (M, R). As Mikulevicius and Rozovskii(1998) pointed out, 39
L2loc (M, R) is not complete. To overcome this we need to introduce the nor-
malized martingale.
Definition. M ∈ H2 (Q0 ) is called an (M, Q0 )-normalized martingale

if there exists a sequence (`n ) ⊂ L2 (M, R) such that
lim kM − I(`n )kH2 (Q0 ) = 0,

n→∞
2
and M ∈ Hloc (Q0 ) is called a local (M, Q0 )-normalized martingale if
there exists a sequence of stopping times (τn ) such that M τn = {Mt∧τn ; t ∈
[0, T0 ]} is a normalized Q0 -martingale for each n.
Remark. Mikulevicius and Rozovskii(1998)39 show that every local

(M, Q0 )-normalized martingale can be viewed as the normalized integral
of a
predictable process with respect to the LSI cylindrical martingale
M = M ` ; ` ∈ R with values in the covariance spaces.
From the representation property of Q0 -martingales, we have the fol-

lowing proposition
Proposition 3.1. Let M be an (M, Q0 )-normalized martingale, then there

exists an Rd -valued predictable process {l1 (u); u ∈ [0, T0 ]} and a P-
e
n 2
measurable function l2 (u, y) and a sequence (` ) ⊂ L (M, R) such that
Z t Z tZ
Mt = M0 +Mtl1 ,l2 := M0 + l1 (u)dW
f (u)+ l2 (u, y){µ(du, dy)−ν(du, dy)}
0 0 E
and
Z Z 2
T0
n 2 n D(u,y,·)
lim EQ0 ù (S(u, ·))−l1 (u) + ù e − 1 − l2 (u, y) ϕ(u, y)λu (dy) du = 0.
n→∞ 0 E
Proposition 3.2. Let M be an (M, Q0 )-normalized martingale, then M is

a martingale under any EMM Q of the discounted bond price maturing at
[T0 , T ∗ ] whose density process Z with respect to Q0 belongs to H2 (Q0 ).
Proof. One proves that ZM is a uniformly integrable martingale under

Q0 and hence M is a uniformly integrable martingale under Q.
Corollary 3.1. Let M l1 ,l2 be a local (M, Q0 )-normalized martingale, then

for any Z z1 ,z2 ∈ ZEnt
Z
0
z1 (u) l1 (u) + ϕ(u, y)z2 (u, y)l2 (u, y)λu (dy) = 0.
E
Proof. For any Z z1 ,z2 ∈ ZEnt , one can see that Z z1 ,z2 can be rewritten as
Ztz1 ,z2 = Zt0 E M z1 ,z2 t ,

where
Z t Z tZ
Mtz1 ,z2 := z1 (u)0 dW
f (u) + z2 (u, y)(µ(du, dy) − ν̃(du, dy)).
0 0 E
Since (z1 , z2 ) satisfies
Z T0 Z T0 Z
2
kz1 (u)k du + ϕ(u, y)z2 (u, y)2 λu (dy)du < ∞, Q0 -a.s.,
0 0 E
z1 ,z2
one can see that M is a locally square integrable martingale under
Q0 , thus there exists a sequence of stopping n o with τn % T0
times (τn )n∈N
l1 ,l2
as n → ∞ such that for each n, Mt∧τn ; t ∈ [0, T0 ] is a (M, Q0 )-
z1 ,z2
normalized martingale and Mt∧τ n
; t ∈ [0, T0 ] is a square integrable mar-
tingale
n under Q0 . Similar o
to the proof of Proposition 3.2, one can show that
l1 ,l2 z1 ,z2
Mt∧τn Mt∧τn ; t ∈ [0, T0 ] is a uniformly integrable martingale under Q0
n o
for each n, thus Mtl1 ,l2 Mtz1 ,z2 ; t ∈ [0, T0 ] is a local martingale under Q0
and
Z

l1 ,l2 Q0
M , M z1 ,z2 t = z1 (u)0 l1 (u) + ϕ(u, y)z2 (u, y)l2 (u, y)λu (dy) = 0.
E
From Theorem 4.2 of De Donno and Pratelli19 , we may adapt the fol-
lowing corollary
Corollary 3.2. If the martingale measure is unique, i.e. there is no
nonzero pair (z1 , z2 ) satisfying (5) for each T ∈ [T0 , T ∗ ], then the mar-
ket is approximately complete, i.e., for every ξ ∈ L2 (Q0 , FT0 ), there is
an (M, Q0 )-normalized martingale M such that MT0 = ξ.
Remark. Similar to De Donno and Pratelli19 , a self-financing portfolio

is an R-valued predictable process π ∈ L2loc (M, R) and the corresponding
wealth process is defined as
Z t
Xtx,π = x+ πu dMu
Z0 t
= x+ πu (S(u, ·))0 dW
f (u)
0
Z tZ
+ πu (eD(u,y,·) − 1)(µ(du, dy) − ν̃(du, dy)),
0 E
then every (M, Q0 )-normalized martingale can be viewed as an approxi-

mate wealth process.
3.3. The Dynamics of the MEM Based on [T0 , T ∗ ]

We now consider Problem (MEM) and we will describe the dynamics of
the MEM based on [T0 , T ∗ ] by a backward semimartingale equation (BSE)
based on a (M, Q0 )-normalized martingale. We need to adapt the optimality
principle (see e.g. Mania, Santacroce and Tevzadze(2003)37) as the following
Lemma 3.3. 1) There exists an RCLL semimartingale denoted by V (t)

such that for each t ∈ [0, T0 ]

V (t) = ess inf E Zt,T0 ln(Zt,T0 )Ft , a.s.,
Z∈ZEnt

which is the largest RCLL process with V (T0 ) = 0 such that Zt {V (t) +
ln(Zt )}; t ∈ [0, T0 ] is a submartingale for each Z ∈ ZEnt .
2) The following three properties are equivalent:
(i) Ẑ is the minimal entropy martingale based on [T0 , T ∗ ];

(ii) Ẑ is optimal for all conditional criteria, i.e., for each t ∈
[0, T0 ]

V (t) = E Ẑt,T0 ln(Ẑt,T0 )Ft , a.s.;
n o
(iii) Ẑt V (t) + ln(Ẑt ) ; t ∈ [0, T0 ] is a uniformly integrable
martingale.
We consider the following BSE
 Z t
∗ 0 0


 a(t) = a(0) + l1 (u) − (φ(u) + θ1 (u)) dW (u)
0


 Z t

 1 2 ∗ 0
+ kφ(u) + θ1 (u)k − l1 (u) φ(u) du


Z0 Z 2



t
l2∗ (u, y) − ln ϕ(u, y) 1 + θ2 (u, y)


 + µ(du, dy)
Z0 t ZE




ϕ(u, y) 1 + θ2 (u, y) − l2∗ (u, y) − 1 ν(du, dy).



 +
0 E



a(T0 ) = 0 .
(17)
A solution of the BSE (17) is a 5-tuple (a; l1∗ , l2∗ ; θ1 , θ2 ) satisfying (17) such
that
(1) l1∗ is an Rd -valued predictable process and l2∗ is a P-measurable

e function
such that
Z t Z tZ
l∗ ,l∗
Mt 1 2
:= l1∗ (u)0 dW
f (u) + l2∗ (u, y){µ(du, dy) − ν̃(du, dy)}
0 0 E
is a local (M, Q0 )-normalized martingale;

(2) θ1 is an Rd -valued predictable process, while θ2 is a P-measurable
e
θ1 ,θ2
function with θ2 (u,y) > −1 such that Z ∈ Z Ent
;
(3) for any Z ∈ ZEnt , Zt {a(t) + ln(Zt )}; t ∈ [0, T0 ] belongs to the class
(D).
Theorem 3.1. Let Assumption 3.1 be satisfied, and assume that the BSE
(17) has a solution denoted by (a; l1∗ , l2∗ ; θ1 , θ2 ), then V (t) = a(t), a.s. for
every t ∈ [0, T0 ] and the MEM based on [T0 , T ∗ ] is given by Z θ1 ,θ2 .
Proof. The idea is the same as the proof of Theorem 3.3 of Kohlmann and
Xiong(2008)33 , but the computation is different. For any pair (z1 , z2 ) such
that
Z z1 ,z2 ∈ ZEnt , by applying Itô’s formula, one can see that
Z t 0
a(t) + ln Ztz1 ,z2 = a(0) + l1∗ (u) + z1 (u) − θ1 (u) dW (u)

Z t 0
1 1 2
+ kφ(u) + θ1 (u)k2 − l1∗ (u)0 φ(u) − φ(u) + z1 (u) du
Z0 t Z 2 2
∗

+ l2 (u, y) − ln 1 + θ2 (u, y) + ln 1 + z2 (u, y) µ(du, dy)
Z0 t ZE
∗
+ ϕ(u, y) θ2 (u, y) − l2 (u, y) − z2 (u, y) ν(du, dy),
0 E
thus
Ztz1 ,z2 {a(t) + ln Ztz1 ,z2 }

Z t
z1 ,z2 z1 ,z2

= a(0) + Zu− a(u−) + ln Zu− (φ(u) + z1 (u))0 dW (u)
0 Z t Z
z1 ,z2 z1 ,z2
+ Zu− a(u−) + ln Zu−
0 E
× ϕ(u, y) 1 + z2 (u, y) − 1 (µ(du, dy) − ν(du, dy))
Z t 0
z1 ,z2
+ Zu− l1∗ (u) + z1 (u) − θ1 (u) dW (u)
Z0 t
z1 ,z2 1 1 2
+ Zu− kφ(u) + θ1 (u)k2 − l1∗ (u)0 φ(u) − φ(u) + z1 (u) du
Z0 t Z 2 2
z1 ,z2
+ Zu− l2∗ (u, y) − ln 1 + θ2 (u, y) + ln 1 + z2 (u, y) µ(du, dy)
Z0t ZE
z1 ,z2
+ Zu− ϕ(u, y) θ2 (u, y) − l2∗ (u, y) − z2 (u, y) ν(du, dy)
0 E
Z t
z1 ,z2
+ Zu− (φ(u) + z1 (u))0 l1∗ (u) + z1 (u) − θ1 (u) du
Z0 t Z
z1 ,z2
+ Zu− ϕ(u, y) 1 + z2 (u, y) − 1
0 E

× l2∗ (u, y) − ln 1 + θ2 (u, y) + ln 1 + z2 (u, y) µ(du, dy)
Z t
z1 ,z2 1
= a(0) + mt + Zu− kθ1 (u) − z1 (u)k2 + z1 (u)0 l1∗ (u) du
0Z 2
Z t
z1 ,z2
+ Zu− ϕ(u, y) z2 (u, y)l2∗ (u, y) + θ2 (u, y) − z2 (u, y)
0 E

+ 1 + z2 (u, y) ln 1 + z2 (u, y) − ln 1 + θ2 (u, y) ν(du, dy)
where
Z t
z1 ,z2 z1 ,z2
mt := Zu− a(u−) + ln Zu− (φ(u) + z1 (u))0
0
0
+ l1∗ (u) + z1 (u) − θ1 (u) dW (u)
Z t Z
z1 ,z2 z1 ,z2
+ Zu− a(u−) + ln Zu− ϕ(u, y) 1 + z2 (u, y) − 1
0 E ∗
+ ϕ(u, y) 1 + z2 (u, y)
l2 (u, y) − ln 1 + θ2 (u, y)

+ ln 1 + z2 (u, y) (µ(du, dy) − ν(du, dy))
∗ ∗
is a martingale under P . Since M l1 ,l2 is a local (M, Q0 )-normalized mar-
tingale, according to Corollary 3.1, one can see that
Z
0 ∗
z1 (u) l1 (u) + ϕ(u, y)z2 (u, y)l2∗ (u, y)λu (dy) = 0,
E
thus
Z t
1
Ztz1 ,z2 {a(t) Ztz1 ,z2 } z1 ,z2
kθ1 (u) − z1 (u)k2 du

+ ln = a(0) + mt + Zu−
Z tZ 2 0
z1 ,z2
+ Zu− ϕ(u, y) θ2 (u, y) − z2 (u, y)
0 E

+ 1 + z2 (u, y) ln 1 + z2 (u, y) − ln 1 + θ2 (u, y) ν(du, dy)
is a local submartingale under P , since for all z2 (u, y) > −1,

θ2 (u, y) − z2(u, y) + (1 + z2(u, y)) ln 1 + z2(u, y) − ln 1 + θ2(u, y) ≥ 0
θ1 ,θ2 θ1 ,θ2
and equality holds when z 2 (u, y) = θ 2 (u, y). Also Z t {a(t)+ln Z t };
t ∈ [0, T0 ] is alocal martingale. By the definition of the solution, for any
Z z1 ,z2 ∈ ZEnt , Ztz1 ,z2 {a(t) + ln(Ztz1 ,z2 )}; t ∈ [0, T0 ] belongs to class (D),
thus Ztl1 ,l2 {a(t)+ln Ztl1 ,l2 }; t ∈ [0, T0 ] is a submartingale. From Lemma

3.3 we thus have

Vt ≥ a(t), a.s.
Furthermore, since Z θ1 ,θ2 ∈ ZEnt and Ztθ1 ,θ2 {a(t)+ln Ztθ1 ,θ2 }; t ∈ [0, T0 ]

is a uniformly integrable martingale,

h i
Ztθ1 ,θ2 a(t) + ln Ztθ1 ,θ2 = E ZTθ1 ,θ2 a(T ) + ln ZTθ1 ,θ2 Ft

= E Z θ1 ,θ2 ln Z θ1 ,θ2 )Ft ,

T T
thus
h i
θ1 ,θ2 θ1 ,θ2
a(t) = E Zt,T ln Zt,T Ft

≥ ess inf Z∈ZEnt E Zt,T0 ln(Zt,T0 )Ft = V (t),
thus a(t) = V (t), a.s., and Z θ1 ,θ2 is the MEM based on [T0 , T ∗ ].
The following theorem improves Theorem 3.5 of Kohlmann and

Xiong(2008)33.
Theorem 3.2. Let Assumption 3.1 be satisfied, then the BSE (17) has a
solution if and only if there exists a Z z1 ,z2 ∈ ZEnt satisfying the reverse
Hölder inequality REnt (P ), which is of the following form
ZTz10,z2 = exp c + MTl10,l2 , a.s.,

where M l1 ,l2 is a local (M, Q0 )-normalized martingale such that

z1 ,z2 l1 ,l2
Zt Mt ; t ∈ [0, T0 ]
is a uniformly integrable martingale.
Similar to the proof of Lemma 3.3 of Delbaen et al (2002)22 , we can

show that the following corollary holds:
Corollary 3.3. Let Assumption 3.1 be satisfied, and assume that the BSE
(17) has a solution denoted by (a; l1∗ , l2∗ ; θ1 , θ2 ), then for any Z z1 ,z2 ∈ ZEnt ,
∗ ∗
Z z1 ,z2 M l1 ,l2
is a uniformly integrable martingale.
When the martingale measure is unique, i.e., there is no nonzero pair

(z1 , z2 ) satisfying (5) for each T ∈ [T0 , T ∗ ], by the boundedness of φ and ϕ,
we see that ln(ZT00 ) ∈ L2 (Q0 , FT0 ), and from Corollary 3.2 that there exists
0 0
an (M, Q0 )-normalized martingale M l1 ,l2 such that
l0 ,l02
MT10 = ln(ZT00 ).
Thus we have the following corollary:
Corollary 3.4. If the martingale measure is unique, set

0

a(t) = EQ0 [ln(Zt,T0
)Ft ],
then (a; l10 , l20 ; 0, 0) is the solution of the BSE (17).
Given n maturities T1 , · · · , Tn with T0 ≤ T1 < T2 < · · · < Tn ≤ T ∗ .

Let ZT1 ,··· ,Tn be the set of strictly positive martingale Z z1 ,z2 on [0, T0 ] with
Z0z1 ,z2 = 1 such that for any maturity Ti ∈ [T0 , T ∗ ], Ztz1 ,z2 p̄(t, Ti ) is a local
martingale on [0, T0 ]. Similarly, we let
 h i 
 z1 ,z2 z ,z z ,z 
 Z ∈ ZT1 ,··· ,Tn satisfies E ZT1 2 ln ZT1 2 < ∞ and 
Z Z T Z 0 0
ZT1 ,··· ,Tn := Z z1 ,z2 T0 0 ,
Ent 
 2
kz1 (u)k du + ϕ(u, y)z2 (u, y) λu (dy)du < ∞, Q -a.s. 
2 0


0 0 E
and consider the minimal entropy martingale (MEM) based on

T1 , · · · , Tn which is defined as the solution of the following optimization
problem:

min E ZT ln ZT .
T ,··· ,Tn
1
Z∈ZEnt
In general, the MEM based on [T0 , T ∗ ] may not be the MEM based on
T1 , · · · , Tn , however, we have the following corollary
Corollary 3.5. Let Assumption 3.1 be satisfied. Assume that the BSE (17)
has a solution denoted by (a; l1∗ , l2∗ ; θ1 , θ2 ). If the local (M, Q0 )-normalized
∗ ∗
martingale M l1 ,l2 satisfies
Z t
l∗ ∗
1 ,l2
Mt = `∗u dM(u),
0
Pn
where `∗u (dT )
= j
and where lj is a predictable process and
j=1 lu δTj (dT ),
δTj (dT ) is the Dirac measure at the point Tj for each j, then Z θ1 ,θ2 is the
minimal entropy martingale based on T1 , · · · , Tn .
This corollary allows to compare the results on bond markets with a finite
number of maturities and the more general situation here.
3.4. An Example
In a situation which should be compared to (2.3), it is possible to give an
example in which we can solve the BSE (17) explicitly. Assume that S(u, T )
is given by
S(u, T ) = s1 (T − u) + s2 (T − u)2
where s1 and s2 are two deterministic Rd -valued vector with s01 s2 = 0, and
eD(u,y,T ) − 1 is given by
eD(u,y,T ) − 1 = d(y)(T − u)
for some deterministic continuous function (y). To simplify the presentation,
we again assume that φ is a deterministic Rd -valued vector and ϕ(y) is
a deterministic bounded strictly positive continuous function on R, while
λu (dy) = λ(du) is a deterministic measure on R which is independent of u,

thus ν(du, dy) = λ(dy)du. Then the density process of Q0 with respect to
P is given by
Z Z
0 0
Zt = E φ W + (ϕ(y) − 1){µ(du, dy) − λ(dy)du} ,
R t
0
and under Q , the discounted price process of the bond maturing at T is
given by
Z t
p̄(u−, T ) s01 (T − u) + s02 (T − u)2 dW

p̄(t, T ) = p̄(0, T ) + f (u)
Z tZ 0
+ p̄(u−, T )d(y)(T − u){µ(du, dy) − ν̃(du, dy)}
0 R
where
W
f (t) = W (t) − φt
ν̃(du, dy) = ϕ(y)λ(dy)du.
The following theorem can then be proved:
Theorem 3.3. If the following deterministic equations
 Z Z
 ks k2 γ + s0 s γ + d(y)e d(y)γ1 λ(dy) = s0 φ + ϕ(y)d(y)λ(dy),
1 1 1 2 2 1
R R
 s0 s γ + ks k2 γ = s0 φ,
1 2 1 2 2 2
(18)
has a solution denoted by (γ1 , γ2 ), let
l1 (u) := γ1 s1 + γ2 s2 ,
 ∗
 2∗ (u, y) := γ1 d(y),


l


θ1 (u) := γ1 s1 + γ2 s2 − φ,
eγ1 d(y)



 θ2 (u, y) :=
 −1
ϕ(y)
Z n
eγ1 d(y) (γ1 d(y) − 1) + 1 λ(dy)
1
kγ1 s1 + γ2 s2 k2 +
o
and a(0) := T0 ×
2 R
and assume that a is given by
Z t
a(t) = a(0) + l1∗ (u)0 − (φ(u) + θ1 (u))0 dW (u)
Z t 0
1
+ kφ(u) + θ1 (u)k2 − l1∗ (u)0 φ(u) du
Z0 t Z 2
∗

+ l2 (u, y) − ln ϕ(u, y) 1 + θ2 (u, y) µ(du, dy)
Z0 t ZE
+ ϕ(u, y) 1 + θ2 (u, y) − l2∗ (u, y) − 1 ν(du, dy),
0 E
then (a; l1∗ , l2∗ ; θ1 , θ2 ) is the solution of the BSE (17) and the minimal entropy
martingale is given by Z θ1 ,θ2 .
Proof. From Itô’s formula applied to ln Ztθ1 ,θ2 and the equations (18), we
have
Z tZ
θ1 ,θ2
ln Zt = (γ1 s1 + γ2 s2 ) W (t) +
0f
γ1 d(y){µ(du, dy) − ν̃(du, dy)}
0 R n
Z
1 γ1 d(y)
ks1 γ1 + γ2 s2 k + (γ1 d(y) − 1)e
o
2
+ + 1 λ(dy) t
2 R
and hence
Z T0 Z
= c + (γ1 s1 + γ2 s2 )0 W γ1 d(y){µ(du, dy) − ν̃(du, dy)},
θ ,θ2
ln ZT10 f (T0 ) +
0 R
where
Z n
1 γ1 d(y)
ks1 γ1 + γ2 s2 k2 + (γ1 d(y) − 1)e
o
c= + 1 λ(dy) × T0 .
2 R
Furthermore, trivially there exists an R-valued function `∗u (dT ) such that
 Z T∗
(T − u)`∗u (dT ),

 γ1 =


T0 ∗ (19)
Z T
 2 ∗
 γ2 = (T − u) ù (dT ),


T0
Z Z
thus (γ1 s1 +γ2 s2 )0 W
f+ γ1 d(y){µ(du, dy)−ν̃(du, dy)} is a local (M, Q0 )-
R
normalized martingale, and from Theorem 3.2 we get that Theorem 3.3
holds.
3.5. Optimal Exponential Utility

In this section, we assume that the BSE (17) has a solution denoted by
(a; l1 , l2 ; θ1 , θ2 ). As an application, we now consider the optimal utility of
an investor with exponential utility
U (x) = − exp{−k0 x},
where k0 is a positive constant.
Definition. An admissible self-financing approximate wealth pro-

cess X with the initial wealth x > 0 is a local (M, Q0 )-normalized martin-
gale with X0 = x such that XZ z1 ,z2 is a uniformly integrable martingale
for any Z z1 ,z2 ∈ ZEnt .
The set of all admissible self-financing approximate wealth processes

with the initial wealth x is denoted by W(x).
We consider the following optimization problem:
Problem (exp-utility opt): sup EU (XT0 ).

X∈W(x)
As U 0 (x) = k0 exp(−k0 x) and I(y) := (U 0 )−1 (y) is given by

ln y − ln k0
I(y) = − ,
k0
similar to Theorem 5.2 of Kohlmann and Xiong(2008)33 , we have the fol-
lowing theorem
Theorem 3.4. Assume that the BSE (17) has a solution denoted by
(a; l1∗ , l2∗ ; θ1 , θ2 ), then the optimal exp-approximate wealth process (i.e., the
solution of Problem (exp-utility opt)) is given by
1 l∗1 ,l∗2
Xt∗ = x − M .
k0 t
Proof. From Theorem 3.2 we get X ∗ ∈ M(x) and from the proof of The-
orem 3.2, we derive that
l∗ ,l∗
ln ZTθ10 ,θ2 = a(0) + MT10 2
.
Let λ∗ := k0 e−k0 x−a(0) , then one can see that
ln(λ∗ ZTθ10 ,θ2 ) − ln k0

I(λ∗ ZTθ10 ,θ2 ) = −
k0
ln(λ∗ ) + a(0) − ln k0 1
=− − MTl10,l2
k0 k0
1
= x − MTl10,l2 = XT∗0 .
k0
Thus for any X ∈ M(x),
E{U (XT0 ) − U (XT∗0 )} ≤ E nU 0 (XT∗0 ) XT0 − XT∗0

o
= E U 0 (I(λ∗ ZTθ10 ,θ2 )) XT0 − XT∗0
n o
= E λ∗ ZTθ10 ,θ2 XT0 − XT∗0
= λ∗ (x − x) = 0,
thus X ∗ is the solution of Problem (exp-utility opt).
From the discussion in the papers on this subject and from the short sum-
mary of this in the introduction it makes no sense to look for an admissible
self-financing portfolio replicating X ∗ . Also the direct valuation problem
of a contingent claim in this setting leads to problems with the intuitive
economic interpretation. Thus we turn to the indifference valuation which
really does make sense.
3.6. The Dynamic Exp Utility Indifference Valuation

We now consider the dynamic exp utility indifference valuation (UIV) for
a given bounded contingent claim H ∈ L∞ (FT0 ). Let W be the set of
all local (M, Q0 )-normalized martingales X such that XZ z1 ,z2 is a uni-
formly integrable martingale for any Z z1 ,z2 ∈ ZEnt , and let W b := {X ∈
W; X is bounded }. Similar to Mania and Schweizer(2005)38, we can de-
fine dynamic exp UIV as the following
Definition. The dynamic exp utility indifference value process for a given
H ∈ L∞ (FT0 ) is a bounded RCLL semimartingale, denoted by C(H) =
{Ct (H); t ∈ [0, T0 ]}, satisfying

ess sup E U (XT0 − Xt )Ft = ess sup E U Ct (H) + XT0 − Xt − H Ft ,
X∈W X∈W
(20)
where U (x) = − exp{−k0 x}.
Remark. If the bond market is approximately complete or H is approxi-

mately hedgeable, (i.e., H = XT , for some X ∈ W), then C(H) = 0.
Let Assumption 3.1 be satisfied and assume that the BSE (17) has a
solution denoted by (a; l1 , l2 ; θ1 , θ2 ), one can define the minimal entropy
martingale measure Q∗ by
dQ∗
= ZTθ10 ,θ2 .
dP FT0
Similar to 38, we have the following lemma from (3.2) and (20)
Lemma 3.4. Let Assumption 3.1 be satisfied and assume that the BSE (17)
has a solution denoted by (a; l1∗ , l2∗ ; θ1 , θ2 ), then the exp utility indifference
value process C(H) can be determined by
ek0 Ct (H) = ess inf EQ∗ e−k0 (XT0 −Xt −H) Ft

X∈W
= ess inf EQ∗ e−k0 (XT0 −Xt −H) Ft .

X∈W b
The following lemma is a version of Girsanov’s theorem
Lemma 3.5. Let Assumption 3.1 be satisfied and assume that the BSE
(17) has a solution denoted by (a; l1∗ , l2∗ ; θ1 , θ2 ). Then
Z t Z t
W ∗ (t) := W
f (t) − θ1 (u)du = W (t) − {φ(u) + θ1 (u)}du
0 0
d ∗
is a R -valued Brownian motion under Q and
ν ∗ (du, dy) := (1 + θ2 (u, y))ν̃(du, dy) = (1 + θ2 (u, y))ϕ(u, y)λu (dy)du
is the compensator of µ under Q∗ .
Remark. For any M l1 ,l2 ∈ W, since M l1 ,l2 is a local (M, Q0 )-normalized

martingale and thus M l1 ,l2 a local martingale under Q∗ , M l1 ,l2 can also be
written as
Z t Z tZ
Mtl1 ,l2 = l1 (u)0 dW ∗ (u) + l2 (u, y)(µ(du, dy) − ν ∗ (du, dy)).
0 0 E
∞
For any H ∈ L (FT0 ), we consider the following BSE
 Z t Z t Z k0 ζ(u,y)
1 2 e −1
 h
 t
 = h 0 − k0 kξ(u)k du − − ζ(u, y) ν ∗ (du, dy)


 Z 0t 2 0 E k0


+ {l̂1 (u) + ξ(u)}0 dW ∗ (u)


 Z0 t Z
+ {l̂2 (u, y) + ζ(u, y)}(µ(du, dy) − ν ∗ (du, dy)),




0 E


hT = B.

(21)
The bounded solution of the BSE (21) is a 5-tuple (h; l̂1 , l̂2 ; ξ, ζ) which
satisfies (21) and
i) l̂1 is a predictable process and l̂2 is a P-measurable

e function such that
l̂1 ,l̂2 b
M ∈W ;
ii) ξ is a predictable process and ζ is a P-measurable
e function such that
for any M l1 ,l2 ∈ W b , (ξ, ζ) is ‘exp-orthogonal’ to (l1 , l2 ) in the
following sense
ek0 ζ(u,y) − 1 ek0 l2 (u,y) − 1

Z
0
ξ(u) l1 (u)+ (1+θ2 (u, y))ϕ(u, y)λu (dy) = 0;
E k0 k0
(note that this is inspired by a construction in Morlais(2008)40
iii) h is a bounded RCLL semimartingale.
Then we get from the standard martingale optimality principle (see e.g.
Proposition 12 of Mania and Schweizer(2005)38):
Theorem 3.5. Let Assumption 3.1 be satisfied, assume that the BSE (17)
has a solution denoted by (a; l1∗ , l2∗ ; θ1 , θ2 ) and the BSE (21) has a bounded
solution (h; l̂1 , l̂2 ; ξ, ζ) for given H ∈ L∞ (FT0 ), then for any t ∈ [0, T0 ],
Ct (H) = ht , a.s.
4. Outlook
It is seen that by leaving the basic somewhat unrealistic model of the first
sections towards more realistic ones many technical and interpretational
difficulties arise. One way out of these deficiencies might be to consider more
structured models like affine term structures. To this end it will be necessary
to study Wishart processes (see Bru(1991)9). There is a lot of research going
on in this field, and some results look rather promising. A different approach
might be to consider markets consisting of a stock and many tradable bonds.
However, for such markets arbitrage and possibly completeness has to be
reconsidered in the sense that the new objects have to be compatible with
the standard market objects to be able to redefine risk-neutral markets.
Some attempts to develop such concepts of “compatitibility” in different
situations are found e.g. in Jacod and Protter(2006)28 and in a manuscript
on the compatible bond-stock market with jumps which we are working on
right now.
References
1. Arai, T. (2005). An extension of mean-variance hedging to the discontinuous
case. Finance Stochast. 9, 129-139 (2005)
2. Baran, M., Jakubowski J., Zabczyk J. (2008): On incompleteness of bond
markets with infinite number of random factors, submitted ; Preprint IMPAN
No. 690; arXiv:0809.2270.
3. Becher, D.(2006). Bounded solution to backward SDE’s with jumps for utility
optimization and indifference hedging, The Annals of Applied Probability,
Vol. 16, No. 4, 2027-2054. DOI: 10.1214/105051606000000475.
4. Biagini, F.(2001): Quadratic Hedging Approaches for Interest Rate Models
with Stochastic Volatility, PhD Thesis, Scuola Normale Superiore.
5. Björk,T.(1997): Interest rate theory in Biais, B. (ed.) et al., Financial mathe-
matics. Lectures given at the 3rd session of the Centro Internazionale Matem-
atico Estivo (CIME), held in Bressanone, Italy, July 8-13, 1996. Berlin:
Springer. Lect. Notes Math. 1656, 53-122 (1997).
6. Björk, T., Di Masi, G., Kabanov, Y., Runggaldier, W.(1997): Towards a
general theory of bond markets. Finance Stoch. 1, 141-174.
7. Björk, T., Kabanov, Y., Runggaldier, W.(1997): Bond market structure in

the presence of marked point processes. Math. Finance 7, 211-223.
8. Bobrovnytska, O. and Schweizer, M. (2004). Mean-variance headging and
stochatic control: Beyond the Brownian setting. IEEE Transactions on Au-
tomatic Control 49, 396-408(2004).
9. Bru, M.F. (1991). Wishart Processes, J.Theoretical Probability 4:4, 725-751.
10. Carmona,R. and Tehranchi, M. (2004): A Characterization of Hedging Port-
folios for Interest Rate Contingent Claims. Ann.Appl.Proba., Vol 14,3, 1267-
1294.
11. Černý, A. and Kallsen, J.(2007). On the structure of general mean-variance
hedging strategies. The Annals of Probability 35,No. 4, 1479-1531(2007).
12. Černý, A. and Kallsen, J.(2008). A counterexample concerning the variance-
optimal martingale measure. Mathematical Finance, Vol. 18, Issue 2, pp.
305-316, April 2008.
13. Choulli, T., Krawczyk,L., and Stricker,C. (1998). E -martingales and their
applications in mathematical finance, The Annals of Probability, 1998, Vol.26,
No.2,853-876(1998).
14. Choulli, T. and Stricker, C.(2005). Minimal entropy-Hellinger martingale
measure in incomplete market. Math. Finance 15, 465-490(2005).
15. Choulli, T. and Stricker, C.(2006). More on minimal entropy-Hellinger mar-
tingale measure. Math. Finance 16, 1-19(2006).
16. Choulli, T, Stricker, C. and Li, J.(2007). Minimal Hellinger martingale mea-
sures of order q. Finance Stoch., Vol. 11, no.3, 399-422(2007).
17. Cont, R.(2004): Modeling term structure dynamics: An infinite dimensional
approach. Int. J. Theor. Appl. Finance 8, No. 3, 357-380 (2005).
18. De Donno, M.(2004): The term structure of interest rates as a random field: A
stochastic integration approach. In Stochastic Processes and Applications to
Mathematical Finance, Proceedings of the Ritsumeikan International Sym-
posium, 2751.
19. De Donno, M. and Pratelli, M. (2004). On the use of measure-valued strate-
gies in bond markets. Finance Stochast. 8, 87-109. DOI: 10.1007/s00780-003-
0102-7.
20. De Donno, M. and Pratelli, M. (2005). A Theory of Stochastic Integration for
Bond Markets. The Annals of Applied Probability, Vol. 15, No. 4, 2773-2791.
21. Delbaen, F. and Schachermayer, W. (1996). The Variance-Optimal Martin-
gale Measure for Continuous Processes. BERNOULLI 2, 81-105(1996).
22. Delbaen, F., Grandits, P., Rheinläder, T., Samperi, D., Schweizer, M. and
Stricker, C. (2002). Exponential hedging and entropic penalties. Math. Fi-
nance, Vol. 12, 99-123.
23. Frittelli, M.(2000). The minimal entropy martingale measure and the valua-
tion problem in incomplete markets. Math. Finance, 10, 39-52.
24. Fujiwara, T. and Miyahara, Y. (2003): The minimal entropy martingale mea-
sures for geometric Levy processes. Finance Stochast. 7, 509-531.
25. Gourieroux, C., Laurent, J.P., Pham, H. (1998): Mean-Variance Hedging and
Numéraire. Mathematical Finance, 8 (3), 179-200.
26. Jacod, J. (1979). Calcul Stochastique et Problèmes de Martingales. Lecture
Notes in Maths. 714. Berlin: Springer.
27. Jacod, J., Shiryaev, A.N. (2003). Limit theorems for stochastic processes.
Berlin Heidelberg NewYork: Springer 2003.
28. Jacod, J., Protter, Ph. (2010). Risk-neutral compatibility with option prices.
to appear in Finance and Stochastics 2010.
29. El Karoui, N. and Quenez, M. C. (1995). Dynamic programming and pricing
of contingent claims in an incomplete market. SIAM J. Control Optim. 33,
29-66(1995).
30. Kohlmann, M. and Tang, S. (2003,a). Multidimensional Backward Stochastic
Riccati Equations and Applications. SIAM Journal on Control and Optimiza-
tion 41, 1696-1721(2003).
31. Kohlmann, M., Xiong, D., and Ye, Z. (2007). Change of filtrations and mean-
variance hedging. Stochastics: An International Journal of Probability and
Stochastic Processes, Vol.79, No. 6, 539-562(2007).
32. Kohlmann, M. and Xiong, D.(2007). The mean-variance hedging of a default-
able option with partial information. Stochastic Analysis and Applications,
25: 869-893(2007).
33. Kohlmann, M. and Xiong, D. (2008). The minimal entropy and the con-
vergence of the p-optimal martingale measures in a general jump model.
Stochastic Analysis and Applications, Volume 26, Issue 5, 941-977(2008).
34. Kohlmann,M., Xiong,D., and Ye,Z.(2010). The mean variance hedging in a
general jump model. Applied Mathematical Finance, 17: 1, 29 57.
35. Xiong, D., Kohlmann, M. (2010). Mean Variance Hedging in a General Jump
Market to appear in Int. Journ. Theory Appl. Finance 2010.
36. Laurent, J.P. and Pham, H. (1999). Dynamic Programming and Mean-
Variance Headging. Finance and Stochastics 3, 83-110.
37. Mania, M., Santacroce M. and Tevzadze, R. (2003). A semimartingale BSDE
related to the minimal entropy martingale measure. Finance Stochast. 7, 385-
402.
38. Mania, M. and Schweizer, M. (2005). Dynamic exponential utility indifference
valuation. The Annals of Applied Probability 2005, Vol. 15, No. 3, 2113-2143.
39. Mikulevicius, R. and Rozovskii, B. L.(1998): Normalized stochastic integrals
in topological vector spaces. In: Séminaire de probabilités XXXII (Lecture
notes in Mathematics). Berlin Heidelberg New York: Springer 1998, 137-165.
40. Morlais, M.A.(2008). Utility maximization in a jump market model. Preprint.
41. Niethammer, C.R. (2007). Portfolio Optimization and Optimal Martingale
Measures in Markets with Jumps. ph.d.thesis University of Konstanz, Ger-
many.
42. Rheinländer, M. and Schweizer, M.(1997). On L2 -projections on a Space of
Stochastic Integrals. Annals of Probability 25, 1810-1831(1997).
43. Schwartz, L.(1994): Semi-martingales banachiques: le theor‘eme des trois op-
erateurs. Seminaire de Probabilites XXVIII (Lect. Notes Math., vol. 1583)
BerlinHeidelbergNew York: Springer 1994.
44. Schweizer, M. (1996). Approximation Pricing and the Variance-Optimal Mar-
tingale Measure.Annals of Probability 24, 206-236(1996).
45. Xiong D. and Kohlmann M. (2010). The mean-variance hedging in a bond
market with jumps. Preprint University of Konstanz, Germany, to appear in
Stochastic Analysis and Applications.
46. Xiong D. and Kohlmann M. (2010). The optimal exponential utility in a

jump bond market. Preprint University of Konstanz, Germany, to appear in
Stochastic Analysis and Applications.
47. Xiong D. and Kohlmann M. (2010). The affine term structure with jumps.
Preprint University of Konstanz, Germany.
48. Xiong D. and Kohlmann M. (2010). The term structure compatible to the
stock price system with jumps. Preprint University of Konstanz, Germany.
May 11, 2011 13:45 WSPC - Proceedings Trim Size: 9in x 6in 08-liu
193
Recombining Tree for Regime-Switching Model: Algorithm and

Weak Convergence
R. H. Liu
University of Dayton
300 College Park, Dayton, OH 45469-2316, USA
Email: ruihua.liu@notes.udayton.edu
This paper is concerned with a multinomial tree method for option pricing
when the underlying asset price follows a regime-switching model. A direct ex-
tension of the well-known CRR tree to the regime-switching model is examined
first. Then an efficient recombining tree is presented that grows linearly as the
number of time steps increases. The weak convergence of the discrete multino-
mial approximations to the continuous regime-switching process is proved via
the method of martingale problem formulation.
Keywords: Regime-switching model; multinomial tree; option pricing; weak

convergence.
1. Introduction
Lattice methods have been broadly adapted in computational finance ever
since the innovative work by Cox, Ross and Rubinstein (Ref. 7) for the
Black-Scholes-Merton option pricing model (the CRR binomial tree). See
Refs. 2,6,8,10,14–16,21–23 and the references therein. From the computa-
tional perspective, a crucial feature for a successful tree design is that the
number of nodes can not grow too fast as the number of time steps increases.
According to Nelson and Ramaswamy (Ref. 21), a binomial approximation
to a diffusion is “computationally simple” if the number of nodes grows at
most linearly in the number of time intervals. One of the main reasons for
the CRR method being practically popular is because it grows linearly and
is therefore computationally simple.
Pricing derivative securities in regime-switching model has drawn con-
siderable attention in recent decade. See Refs. 1,3–5,9,11–13,17,20,24,
among others. In this setting, asset prices are dictated by a number of
194 R. H. Liu
stochastic differential equations coupled by a finite-state Markov chain,

which represents random switch among different regimes. Model parame-
ters (drift and volatility coefficients) are assumed to depend on the Markov
chain. In certain cases, closed-form solutions can be obtained for option
prices. For instance, Guo (Ref. 11) provided an analytical formula for the
European call option prices assuming there are two regimes; Buffington
and Elliott (Ref. 5) provided a general formula for m ≥ 2 regimes; Guo
and Zhang (Ref. 12) studied a perpetual American put option with two
regimes and derived an analytical solution. However, it is extremely diffi-
cult to obtain a closed-form solution for many other cases, in particular,
the American options with finite expiration time and with m ≥ 2 regimes.
Hence seeking efficient numerical approximation schemes comes up as an
important alternative.
In Ref. 19, the author has developed a multinomial tree approach for
regime-switching models which grows linearly as the number of time steps
increases. This recombining tree allows us to use large number of time steps
to obtain accurate approximations for both European and American option
prices. We have also developed an approximation regime-switching model
for the well-known Heston’s stochastic volatility model and then employed
the tree approach for pricing options. It is noted (see, for example Ref. 28)
that using a regime-switching model makes it possible to describe stochastic
volatility in a relatively simple manner (simpler than the so-called stochastic
volatility models). Various numerical examples are studied in Ref. 19. See
Ref. 19 for those numerical results and further discussions of the method.
The present paper is concerned with the weak convergence of the multi-
nomial tree approximations to the continuous regime-switching diffusion
process. For models without regime-switching, results on convergence of
the binomial tree approximations have been obtained in a number of arti-
cles. See Refs. 6,14,15,21,23 and the references therein. However, due to the
added Markov chain in the underlying model, the methods used in these
papers can not be directly applied to our regime-switching case. Instead,
we employ the martingale problem formulation procedure (Ref. 25) to es-
tablish our convergence result. This method is developed by Yin, Song and
Zhang in Ref. 25 for a class of discrete numerical approximations for dif-
fusions with regime-switching. Also see Ref. 28 for detailed accounts and
recent generalizations of the method.
The rest of the paper is organized as follows. Section 2 is concerned
with the tree construction. The regime-switching model is introduced in
Subsection 2.1. A direct extension of the CRR tree to the regime-switching
Recombining Tree for Regime-Switching Model 195
model is examined in subsection 2.2 which is seen not computationally

simple. Our new recombining tree is presented in subsection 2.3. Section 3
is devoted to the weak convergence of the tree approximations. We establish
the convergence result under a general framework that includes both the
completely recombining tree and the partially recombining CRR tree as
special cases. Section 4 provides further remarks and concludes the paper.
2. Regime-Switching Trees and Option Pricing

2.1. Regime-Switching Model
Let αt be a continuous-time Markov chain taking values among m0 different
states, where m0 is the total number of states considered for the economy.
Each state represents a particular regime and is labeled by an integer i be-
tween 1 and m0 . Hence the state space of αt is given by M := {1, . . . , m0 }.
For example, if m0 = 2 (two regimes), then αt = 1 can indicate a bullish
market and αt = 2 a bearish market. Let matrix Q = (qij )m0 ×m0 denote
the generator of αt . From Markov chain theory (see for example, Yin and
Zhang (Ref. 26)), the entries qij in matrix Q satisfy: (I) qij ≥ 0 if i 6= j;
P
(II) qii ≤ 0 and qii = − j6=i qij for each i = 1, . . . , m0 .
We assume that the risk-neutral probability space (Ω, F , P) e is given.
Let Bt be a standard Brownian motion defined on (Ω, F , P) e and assume it
is independent of the Markov chain αt . We consider the following regime-
switching geometric Brownian motion for the risk-neutral process of the
underlying asset price St :
dSt
= (rαt − dαt )dt + σαt dBt , t ≥ 0, (1)
St
where rαt is the instantaneous risk-free interest rate, dαt and σαt are div-
idend rate and volatility rate of the asset, respectively. Note that those
parameters depend on the Markov chain αt , indicating that they can take
different values in different regimes. We assume that σi > 0, for each i ∈ M.
2.2. A Direct Extension of CRR Tree to Regime-Switching

A direct extension of the CRR tree (Ref. 7) to the regime-switching model
(1) proceeds as follows. See Refs. 1,17.
Let T > 0 denote the maturity of option under consideration. Di-
vide the interval [0, T ] into N steps. Thus the time step size is given
T
by h = N . Consider the joint Markov process (St , αt ), 0 ≤ t ≤ T
and let (Sk , αk ) := (St , αt )t=kh be the state at the kth step of the tree,
196 R. H. Liu
k = 0, 1, . . . , N . Assuming (Sk , αk ) = (S, i). Then, at the (k + 1)th step,

the asset price Sk+1 can either move up to Sui with probability pi or move
down to Su−1 i with probability 1 − pi , where
√
σi
√
h eri h − e−σi h
ui = e , pi = √ √ .
eσi h − e−σi h
On the other hand, the Markov chain αk+1 at the (k +1)th step may stay at
state i with probability pα
ii or jump to any other state j 6= i with probability
pα
ij , where the one-step transition probabilities pα
ij of the Markov chain αk
are defined by
pα

ij = P {αk+1 = j αk = i}, 1 ≤ i, j ≤ m0 . (2)

It then follows that at the (k + 1)th step, there are totally 2m0 possible
states for (Sk+1 , αk+1 ), given by
(Sui , i), with probability pi pα

ii ,


 (Su−1 , i), with probability

(1 − pi )pα
i ii ,
(Sk+1 , αk+1 ) = α (3)

 (Sui , j), with probability pi pij , j 6= i,
(Su−1 (1 − pi )pα

i , j), with probability ij , j 6= i.
An illustrative two-regime tree is depicted in Fig. 1.

`

`

`
`H`
H``
```
H `
```

HH ``

`
```
H
H
H
H `` H
HH `
``H
``H
`
` H`
`
H``` HH

HH ``` H
HH ```
H
``
H``` HH
`
HH ```
HH H`
HH
HH `````

HH``
H
H``` H
HH ```HH
HH ```H
`
H`
HH
HH
H
H`
Fig. 1. A non-completely recombining tree for two regimes

It can be seen that the state space for the two-dimensional Markov chain
{(Sk , αk )}k≥0 is the set
n
2l 0 −km0
(S0 u2l 1 −k1 2l2 −k2
u2 · · · umm , i)0 ≤ li ≤ ki , 1 ≤ i ≤ m0 ,

1 0
m0
X o (4)
ki = k, 0 ≤ k ≤ N ,
i=1
where S0 is the initial stock price. It was shown in Aingworth et al. [1,
Proposition
the number of distinct asset prices at the kth period
1] that
k + 2m0 − 1
is = O(k 2m0 −1 ). Thus the tree does not grow linearly and
2m0 − 1
therefore is not computationally simple according to Ref. 21. Consequently,
it works well for small number of regimes and for moderate number of tree
steps. However, if m0 and N are large, such an implementation will become
computationally formidable due to node explosion. For example, consider a
model with 20 regimes.
If 100 time steps were used, then the last step of the
139
tree would have ≈ 4.7 × 10192 nodes. Apparently. most computers
39
can not handle such a heavy computation requirement.
2.3. A Regime-Switching Recombining Tree

To have a tree structure that grows linearly, complete recombination of
nodes needs to be achieved at each time step. For the generalized CRR
tree discussed in the previous section, node recombination is only achieved
partially, namely, nodes do combine if the regime stays the same. However,
whenever a switch of regime occurs, the change of asset price will very likely
produce new nodes, due to different change factor ui in different regime.
To resolve this issue, we adjust the change factor ui in a suitable way.
Meanwhile, it is necessary to match the local mean and variance calculated
from the tree to that implied by the continuous regime-switching diffusion,
in order for the discrete tree approximations to converge to its underlying
continuous process (see next section). We present the tree design in this
section. For numerical experiments and an application of the tree method
to the Heston’s stochastic volatility model, we refer the reader to Ref. 19.
First, let Xt = ln(St /S0 ). Then St = S0 eXt where Xt is the solution of
the stochastic differential equation:
dXt = aαt dt + σαt dBt , X0 = 0, (5)
where
1
aαt = rαt − dαt − σα2 t . (6)
2
198 R. H. Liu
We design a completely recombined tree for the process Xt instead of the

asset price St as done in Refs. 1,3. The corresponding approximations of
asset price can be easily obtained from the approximations of Xt . Divide
T
the option life [0, T ] into N steps and let h = N be the step size. Let
(Xk , αk ) := (Xt , αt )t=kh be the state at the kth step of the tree, k =
0, 1, . . . , N . We use three branches for each regime, a up move, a down
move, and a middle stay (no move). The up and down moves are carefully
chosen so that they must take values among 2b + 1 evenly spaced points,
where the specification of b will be discussed shortly. Let constant σ̄ > 0
and positive integer b be given. Assume Xk = x, then at the next step,
Xk+1 must √ take the three values from the set of 2b + 1 values given by
{x + j σ̄ h, j = −b, −b + 1, . . . , 0, . . . , b − 1, b}. Specifically, for regime i at
step k, i.e., (Xk , αk ) = (x, i), let li (to be determined) be the number of
upward moves of Xk+1√. Then, the three √ branches associated with regime
i are given by x + li σ̄ h, x and x − li σ̄ h. Next, let pi,u , pi,m and pi,d
be the conditional probabilities corresponding to the up, middle and down
branches. By matching the mean and variance implied by the trinomial
lattice to that implied by the SDE (5), we have,
 √ √
 (li√
 σ̄ h)pi,u − (li σ̄√ h)pi,d = ai h,
(l σ̄ h)2 pi,u + (li σ̄ h)2 pi,d = σi2 h + a2i h2 , (7)
 i
 pi,u + pi,m + pi,d = 1,
1
where ai = ri − di − σi2 . It follows that,
2
√
σi2 + ai (li σ̄) h + a2i h
pi,u = 2
,
2(li σ̄)
√
2 2
σ − ai (li σ̄) h + ai h
pi,d = i , (8)
2(li σ̄)2
2 2
σ + ai h
pi,m = 1− i .
(li σ̄)2
Consequently, emanating from the node (x, i), there are 3m0 states for
(Xk+1 , αk+1 ), given by
 √
α
 (x + li σ̄ h, j), with probability pij pi,u ,

(Xk+1 , αk+1 ) = (x, j) √ with probability pαij pi,m , j = 1, . . . m0 ,
 (x − l σ̄ h, j), with probability pα p ,

i ij i,d
(9)
where pα ij denote the one-step transition probabilities of the Markov chain
αk from state i to j (see (2)). It can be seen that the number of nodes
at the kth step is m0 (2bk + 1), linear in k. Hence, our design does yield
a computationally simple tree. If we consider again the previous example,
namely, m0 = 20 and N = 100, then the last step of the new tree will have
20(200b + 1) = 80020 if b = 20 is chosen, a much smaller number comparing
to 4.7 × 10192 .
Fig. 2 displays a recombining tree with 7 branches (b = 3) emanating
from each node (a heptanomial tree structure).
`

`

(((
`

((
((
(

(
((

`
(

h( ((
((
Phhh `
HP H
h
P
P h
((
h

(
(
h
H(
(
(
P h
h
hh(
`
(

h( H P
P h`
P Hh
Ph
H
h
P

H

hH
h(
h
P

(
(((
(
PP

P(
P(
Hh

((((`(

h

(H(
(H Hh
P
P
h
H(
P
hh
P(
h`
P
P
Hh
Ph
hH P
H((

((( HP

hH
h(
h
((
PP

H
((
HP((Hhh
P
((

( ((
P H
P
hhh
H
PH
`(

h
P
Hh
Phhh `P(
(

h
Hh
P
H
h
h
P
H

H(

P(
H(
h`
P
((

HPP hhh HP

PhH
h
((
h
(
hPP

H
H P hhh
H(( H hh
HHPP hh`(

h(
( P
HPH
P P
H(

H
hh
P(
h`
P
H
P
HhhH ((
HHPPP P
H
P
h

h
P(H
h
((
h

P

((
P
Hh
H
P

H

HH PPP` H(
( P HhhPH
hh
P(
(

h
P
(
Hh
Ph
H

hHP
P
PH(((P`
H
h
HH H
P

h(
H
h(
h((
PH
PH
HP
((( Hhh P
HH`(

h(
Ph HPPPHHhhh
Hh
PH`
P
H HH PPH
Hh
PPh
P
hhh
Hh PH
HHPP Hhhh
PP
HP
HH hh
HHPPP H`
HH PH
PH
PP
H`
HH
HH
H`
Fig. 2. A two-step recombining heptanomial tree
Note that if li and h are not chosen properly, it is possible that (8) results
in negative branch probabilities. The following proposition gives conditions
under which the branch probabilities are guaranteed to be non-negative.
Proposition 1. The following assertions hold for the solutions given by

(8).
200 R. H. Liu
(1) If ai = 0, then 0 ≤ pi,u , pi,m , pi,d ≤ 1 provided that li σ̄ ≥ σi ;

(2) If ai 6= 0, then 0 ≤ pi,u , pi,m , pi,d ≤ 1 provided that
√ (a) σi < li σ̄ ≤ 2σi
(li σ̄)2 −σi2 [li σ̄− (li σ̄)2 −4σi2 ]2
and h ≤ a2i
, or (b) li σ̄ > 2σi and h ≤ 4a2i
.
Remark 1. The numerical results in Ref. 19 show that the approximate

option prices are insensitive to the choice of parameter σ̄. Hence the choice
of li can be flexible as long as the conditions in Proposition 1 are satisfied.
One way is to choose the li value so that the upper constraint for h takes
the larger one. Specifically, for given σ̄, let k1 = b 2σ 2σi
σ̄ c and k2 = d σ̄ e be
i
the floor and ceiling of 2σσ̄ , respectively. Then,

i
• if either k1 = k2 or k1 σ̄ < σi√ , then set li = k2 ; otherwise,

(k1 σ̄)2 −σi2 [k2 σ̄− (k2 σ̄)2 −4σi2 ]2
• let a = a2i
,b= 4a2i
, if a ≤ b, then set li = k2 ; if
a > b, then set li = k1 .
Using the recombining tree we have constructed for the underlying asset
process Xt , both European and American options can be priced following
a similar procedure that is used in the CRR tree for valuing options in the
Black-Scholes model, i.e., by starting at the last step of the tree (n = N )
and working backward iteratively. At each step and for each node, we first
calculate the probability weighted average of all option prices at the nodes
in the next step that are directly emanated from the current node. We then
discount the averaged future price using the interest rate of the current
node. This gives us the discounted expectation of the future option value,
where the expectation is taken with respect to the risk-neutral probability
Pe . For American options, in addition we need to check the early exer-
cise possibility by comparing this value with the immediate exercise value.
Because of the presence of regime-switching, those computations become
more involved and special care needs to take. Note that the option values
are functions of the underlying asset price and also depend on the regime.
Consider a put option, let P n (x, i) denote the option value at step n for
the node associated with the state (Xn , αn ) = (x, i). Then, for the terminal
step n = N ,
P N (x, i) = max{K − S0 ex , 0}, i = 1, . . . , m0 , (10)
where K is the strike price of the option. At step n < N , for European put,
m0
X h √
P n (x, i) = e−ri h pα
ij P
n+1
(x + li σ̄ h, j)pi,u
j=1 (11)
n+1
√ n+1
i
+P (x − li σ̄ h, j)pi,d + P (x, j)pi,m ,
and for American put,

m0
(
n x −ri h
h X √
P (x, i) = max K − S0 e , e pα
ij P
n+1
(x + li σ̄ h, j)pi,u
j=1 ) (12)
n+1
√ n+1
i
+P (x − li σ̄ h, j)pi,d + P (x, j)pi,m .
Formulae for call options are obtained by changing K − S0 ex to S0 ex − K

in (10)-(12).
3. Convergence of the Tree Approximations

In this section we establish the convergence results of the discrete approxi-
mation processes implied by the regime-switching trees to the continuous-
time processes. We prove the convergence under a general framework that
includes the completely recombining tree and the partially recombining
CRR tree as two special cases.
We consider the following one-dimensional stochastic differential equa-
tion modulated by the Markov chain αt ,
dXt = µ(Xt , αt )dt + σ(Xt , αt )dBt , t ≥ 0, (13)
where µ(·, ·) : R×M → R and σ(·, ·) : R×M → R are appropriate functions
satisfying the Lipschitz condition
|µ(x1 , i) − µ(x2 , i)| ≤ K|x1 − x2 |, |σ(x1 , i) − σ(x2 , i)| ≤ K|x1 − x2 |, (14)
and the linear growth condition,
|µ(x, i)| ≤ K(1 + |x|), |σ(x, i)| ≤ K(1 + |x|), (15)
for all x, x1 , x2 ∈ R and for each i ∈ M, where K is a positive constant
(In what follows we will use K as a generic positive constant; that is, its
values may vary at different places). In addition, we assume σ(x, i) > 0 for
all x ∈ R and for each i ∈ M. Note that for the regime-switching model
(5) on which the tree construction is based, µ(x, i) = ri − di − 21 σi2 and
σ(x, i) = σi . Clearly, (14) and (15) are satisfied.
For each i ∈ M, let function V (x, i) be twice continuously differentiable
in x. Define an operator L by
dV 1 d2 V
LV (x, i) = µ(x, i) (x, i) + σ 2 (x, i) 2 (x, i) + QV (x, ·)(i) (16)
dx 2 dx
where
X m0
X
QV (x, ·)(i) = qij [V (x, j) − V (x, i)] = qij V (x, j) (17)
j6=i j=1
202 R. H. Liu
and Q = (qij )m0 ×m0 is the generator of αt .

In what follows we will use ε > 0 (instead of h as in the previous
section) for the time step size. Then N = b Tε c is the number of steps for a
tree construction. Let the initial state X0 and initial regime α0 be given.
For each fixed ε > 0, we construct a continuous-time process (X ε (t), αε (t)),
t ≥ 0 such that,
X ε (0) = X0 , αε (0) = α0 , (18)
X ε (t) = Xn , αε (t) = αn , for nε ≤ t < (n + 1)ε, n = 0, 1, . . . , N − 1, (19)

where for given (Xn , αn ) = (x, i), (Xn+1 , αn+1 ) is determined as following:
First for the discrete Markov chain {αn , n ≥ 0}, the transition probability
matrix is given by
P ε = I + εQ, (20)
that implies
P {αn+1 = j|αn = i} = δij + εqij , 1 ≤ i, j ≤ m0 , 0 ≤ n ≤ N − 1, (21)
where δij = 1 if i = j and δij = 0 if i 6= j. It can be shown by applying
Theorem 3.1 in Ref. 27 that the interpolated process αε (·) converges weakly
to the continuous Markov process α(·) generated by Q.
Next, we assume that for given (Xn , αn ) = (x, i), Xn+1 can take n0
different values X ε,l (x, i), l = 1, . . . , n0 with conditional probabilities,
n o
P Xn+1 = X ε,l (x, i)Xn = x, αn = i = pε,l (x, i), l = 1, 2, . . . n0 , (22)

where functions X ε,l (·, ·) : R × M → R satisfy −∞ < X ε,l (x, i) < ∞ for all
x ∈ R and each i ∈ M, and pε,l (·, ·) : R × M → R satisfy
n0
X
0 ≤ pε,l (x, i) ≤ 1, pε,l (x, i) = 1, for all x ∈ R, i ∈ M. (23)
l=1
Moreover, we assume throughout this section that, for i ∈ M, x ∈ R,

h i X n0
E Xn+1 Xn = x, αn = i = pε,l (x, i)X ε,l (x, i) = x + µ(x, i)ε, (24)

l=1
and
n0
h i X 2
E (Xn+1 − x)2 Xn = x, αn = i = pε,l (x, i) X ε,l (x, i) − x

l=1
(25)
2 2 2
= σ (x, i)ε + µ (x, i)ε .
Note that these conditions are all satisfied by the regime-switching recom-
bining tree design presented in Section 2 with n0 = 3.
The following lemma presents an upper bound of the second moments
of Xn , 0 ≤ n ≤ b Tε c which will be used in the weak convergence proof.
Lemma 1. Under Assumptions (14), (15), (24) and (25), for any fixed
T > 0 and any ε > 0, we have
2
sup E [Xn ] ≤ (X02 + KT )eKT < ∞. (26)
0≤n≤bT /εc
Proof. Using (24) and (25) we have,

h i h i
2 2
E Xn+1 Xn , αn = E (Xn + (Xn+1 − Xn )) Xn , αn

= Xn2 + 2Xn µ (Xn , αn ) ε + σ 2 (Xn , αn ) ε + µ2 (Xn , αn ) ε2 .

Using (15) we have,
µ2 (Xn , αn ) ≤ K 1 + Xn2 and σ 2 (Xn , αn ) ≤ K 1 + Xn2 .

It follows that
h i
2
Xn , αn ≤ Xn2 + Kε 1 + Xn2 .

E Xn+1 (27)

Taking the expectation on both sides of (27) leads to

2
≤ E Xn2 + Kε + KεE Xn2 .

E Xn+1 (28)
Iterating (28), we obtain
n−1
X
Xn2 X02 E Xj2 .

E ≤ + Knε + Kε (29)
j=0
Applying the Gronwall’s inequality yields

E Xn2 ≤ (X02 + Knε)eKT ≤ (X02 + KT )eKT ,

and then (26) follows.

It follows immediately that
2
sup E [X ε (t)] ≤ K < ∞. (30)
0≤t≤T
Next, we establish the tightness of the continuous processes {X ε (t) :

0 ≤ t ≤ T } in a suitable function space. To this end, we rewrite the ap-
proximation (22) into the following iterative equation:
Xn+1 = Xn + µ (Xn , αn ) ε + ξ (Xn , αn ) , (31)
204 R. H. Liu
where the random functions ξ(·, ·) are defined as following: For each x ∈ R
and each i ∈ M, ξ(x, i) takes values from the set of n0 values {X ε,l (x, i) −
x − µ(x, i)ε, l = 1, . . . , n0 } with probabilities pε,l (x, i), l = 1, 2, . . . n0 , as
defined in (22) and (23). It follows from (24) and (25) that
h i
E ξ (Xn , αn ) Xn = x, αn = i = 0,

h i (32)
E ξ 2 (Xn , αn ) Xn = x, αn = i = σ 2 (x, i)ε.

Lemma 2. Under Assumptions (14), (15), (24) and (25), the interpolation
process {X ε (t) : 0 ≤ t ≤ T } is tight in D[0, T ], the space of functions
that are right continuous and have left limits, endowed with the Skorohod
topology.
Proof. For any δ > 0, t > 0, 0 ≤ s ≤ δ and t + s ≤ T , we have,

 2
(t+s)/ε−1
2
X
E [X ε (t + s) − X ε (t)] = E  (Xk+1 − Xk )
k=t/ε
 2
(t+s)/ε−1
X
= E (µ (Xk , αk ) ε + ξ (Xk , αk ))
k=t/ε
 2
(t+s)/ε−1
X
≤ 2ε2 E  µ (Xk , αk )
k=t/ε
 2
(t+s)/ε−1
X
+2E  ξ (Xk , αk ) .
k=t/ε
Note that for simplicity, in the above we have used t/ε and (t + s)/ε for the
integer parts of t/ε and (t+s)/ε, respectively. We will adapt this convention
for the rest of the paper.
Note that
 2
(t+s)/ε−1 (t+s)/ε−1
X X
2
E µ2 (Xk , αk )

ε E µ (Xk , αk ) ≤ εδ
k=t/ε k=t/ε
(t+s)/ε−1
X
1 + E[Xk2 ] ≤ Kδ 2 ,

≤ εδK
k=t/ε
and
 2
(t+s)/ε−1
X
E ξ (Xk , αk )
k=t/ε
(t+s)/ε−1
X X
E ξ 2 (Xk , αk ) + 2

= E [ξ (Xk , αk ) ξ (Xl , αl )]
k=t/ε k<l
(t+s)/ε−1
X X
E Ek ξ 2 (Xk , αk ) + 2

= E [ξ (Xk , αk ) El [ξ (Xl , αl )]]
k=t/ε k<l
(t+s)/ε−1 (t+s)/ε−1
X X
E σ 2 (Xk , αk ) ≤ Kε 1 + E[Xk2 ] ≤ Kδ,

=ε
k=t/ε k=t/ε
where we have used Ek for the conditional expectation given (Xk , αk ). We

have
2
E [X ε (t + s) − X ε (t)] ≤ K(δ + δ 2 ).
Then it follows that
2
lim lim sup E [X ε (t + s) − X ε (t)] = 0.
δ→0 ε→0
Then the tightness criterion (see Ref. 18, P47) implies that {X ε (·)} is tight
in D[0, T ].
Note that it can be shown by applying Theorem 3.1 in Ref. 27 that the
interpolated process {αε (·)} is tight. It then follows that {X ε (·), αε (·)} is
tight in D([0, T ]; R × M), the space of functions that are defined on D[0, T ]
taking values in R × M, and that are right continuous and have left limits,
endowed with the Skorohod topology.
Since {X ε (·), αε (·)} is tight, by the Prohorov’s Theorem, it has con-
vergent subsequences. We select such a subsequence and still denote it by
{X ε (·), αε (·)} for notational simplicity. Denote the limit by {X(·), α(·)}.
By Skorohod representation, we may assume without loss of generality that
the subsequence {X ε (·), αε (·)} converges to {X(·), α(·)} w.p.1, and that the
convergence is uniform on any bounded interval. Next we apply the Mar-
tingale problem formulation method (Refs. 25,28) to characterize the limit
process {X(·), α(·)}.
Theorem 1. Under Assumptions (14), (15), (24) and (25), the interpo-
lation process (X ε (·), αε (·)) converges weakly to (X(·), α(·)) as ε → 0 such
that (X(·), α(·)) is the solution of the martingale problem associated with
operator L defined by (16).
206 R. H. Liu
Proof. We are to show that for each i ∈ M and for any function ϕ(·, i) ∈
C02 ,
Z t
Mϕ (t) := ϕ(X(t), α(t)) − ϕ(X0 , α0 ) − Lϕ(X(s), α(s)) ds (33)
0
is a martingale. This is done via a series of approximations that average
out the unwanted terms.
Following Refs. 25,28, the claim (33) will be verified if we can show that
for each i ∈ M, any bounded and continuous function hj (·, i), any positive
integer κ, any 0 < tj ≤ t with j ≤ κ, s > 0, and t + s ≤ T ,
( κ
Y h
E hj (X(tj ), α(tj )) ϕ(X(t + s), α(t + s)) − ϕ(X(t), α(t))
j=1
Z t+s i
) (34)
− Lϕ(X(u), α(u)) du = 0.
t
In what follows we show that (34) holds for the pair (X ε (·), αε (·)) and hence
for the pair (X(·), α(·)). The proof is divided into four steps.
Step 1. We begin by choosing a sequence of positive integers {mε } satisfying
mε → ∞ as ε → 0 but ∆ε := εmε → 0. Again, we use t/∆ε and (t + s)/∆ε
for their integer parts (to avoid the use of the floor function notation).
Introduce the index set
χεl = {k : lmε ≤ k ≤ (l + 1)mε − 1}. (35)
Then we can rewrite ϕ(X ε (t + s), αε (t + s)) − ϕ(X ε (t), αε (t)) as
ϕ(X ε (t + s), αε (t + s)) − ϕ(X ε (t), αε (t))
(t+s)/∆ε −1
X
= ϕ(X(l+1)mε , α(l+1)mε ) − ϕ(X(l+1)mε , αlmε )
l=t/∆ε (36)
(t+s)/∆ε −1
X
+ ϕ(X(l+1)mε , αlmε ) − ϕ(Xlmε , αlmε ) .
l=t/∆ε
Step 2. We work on the second summation (the third line) of (36) in this
step. Noting that
(t+s)/∆ε −1
X
ϕ(X(l+1)mε , αlmε ) − ϕ(Xlmε , αlmε )
l=t/∆ε
(t+s)/∆ε −1
X X
= [ϕ(Xk+1 , αlmε ) − ϕ(Xk , αlmε )] ,
l=t/∆ε k:k∈χεl
we have ,
 
Yκ (t+s)/∆ε −1
ε ε
X 
lim E hj (X (tj ), α (tj )) ϕ(X(l+1)mε , αlmε ) − ϕ(Xlmε , αlmε )
ε→0  
j=1 l=t/∆ε
 
Yκ (t+s)/∆ε −1
X X
ε ε
= lim E hj (X (tj ), α (tj ))  ϕ0x (Xk , αlmε )µ(Xk , αk )ε
ε→0
k:k∈χε

j=1 l=t/∆ε l
(t+s)/∆ε −1
X X
+ ϕ0x (Xk , αlmε )ξ(Xk , αk )
l=t/∆ε k:k∈χε
l
(t+s)/∆ε −1
1 X X
2 2
+ ϕ00
xx (Xk , αlmε )µ (Xk , αk )ε
2
l
(t+s)/∆ε −1
1 X X
2
+ ϕ00
xx (Xk , αlmε )ξ (Xk , αk )
2
l
(t+s)/∆ε −1
X X
+ ϕ00
xx (Xk , αlmε )µ(Xk , αk )ξ(Xk , αk )ε
l

(t+s)/∆ε −1 1
X X Z h  i 2
+ ϕ00 g 00
xx (Xk + u∆xk , αlmε ) − ϕxx (Xk , αlmε ) ∆xk du
g  ,
l=t/∆ε k:k∈χε 0 
l
(37)
gk = µ(Xk , αk )ε + ξ(Xk , αk ). Since ϕ00 (·, i) is continuous with

where ∆x xx
compact support, the weak convergence of (X ε (·), αε (·)), the Skorohod rep-
resentation, and the boundedness of hj (·, i) then yield
κ (t+s)/∆ε −1
(
Y X X Z 1 h
ε ε
lim E hj (X (tj ), α (tj )) ϕ00xx (Xk + u∆x
gk , αlmε )
ε→0 0
j=1 l=t/∆ε k:k∈χεl
)
i 2
− ϕ00xx (Xk , αlmε ) gk du
∆x = 0.
(38)
208 R. H. Liu
Since ϕ0x (·, i) and µ(·, i) are continuous for each i ∈ M, we have
(t+s)/∆ε −1
( κ
)
Y ε ε
X X
lim E hj (X (tj ), α (tj )) ϕ0x (Xk , αlmε )µ(Xk , αk )ε
ε→0
j=1 l=t/∆ε k:k∈χε
l
(t+s)/∆ε −1
( κ
)
Y ε ε
X X
= lim E hj (X (tj ), α (tj )) ε ϕ0x (Xlmε , αlmε )µ(Xlmε , αk )
ε→0
j=1 l=t/∆ε k:k∈χε
l
(t+s)/∆ε −1
( κ
Y X X
= lim E hj (X ε (tj ), αε (tj )) ε Elmε
ε→0
j=1 k:k∈χl ε
" ml=t/∆ε #)
X 0
0
× ϕx (Xlmε , αlmε )µ(Xlmε , i)I{αk =i} .
i=1
Note that when ε → 0 and lεmε = l∆ε → u, (l + 1)∆ε → u and for

any k satisfying lmε ≤ k < (l + 1)mε , εk → u. In addition, in view of the
weak convergence of αε (·) to α(·) and the Skorohod representation (without
changing notation), we may assume that I{αε (u)=i} → I{α(u)=i} w.p.1. As
a result, for the above limit, we have,
κ (t+s)/∆ε −1
(
Y X X
ε ε
lim E hj (X (tj ), α (tj )) ε Elmε
ε→0
"m #)
X 0
× ϕ0x (Xlmε , αlmε )µ(Xlmε , i)I{αk =i}

i=1

Yκ m0 (t+s)/∆
X X ε −1
= lim E hj (X ε (tj ), αε (tj )) ∆ε ϕ0x (X ε (l∆ε ), αε (l∆ε ))
ε→0 
j=1 i=1 l=t/∆ε 
1 X 
µ(X ε (l∆ε ), i) × Elmε  I{αε (εk)=i} 
mε ε

 k:k∈χl 
Yκ m0 Z t+s
X 
=E hj (X(tj ), α(tj )) ϕ0x (X(u), α(u))µ(X(u), i)I{α(u)=i} du
 t 
j=1 i=1

Yκ Z t+s 
=E hj (X(tj ), α(tj )) ϕ0x (X(u), α(u))µ(X(u), α(u)) du .

j=1 t 
(39)
In view of (32), we have

 
Yκ (t+s)/∆ε −1 
X X
E hj (X ε (tj ), αε (tj )) ϕ0x (Xk , αlmε )ξ(Xk , αk )
 ε

j=1 l=t/∆ε k:k∈χl
 
Yκ (t+s)/∆ε −1 
X X
=E hj (X ε (tj ), αε (tj )) ϕ0x (Xk , αlmε )Ek [ξ(Xk , αk )]
 
= 0,
(40)
and similarly,
κ (t+s)/∆ε −1
(
Y X X
ε ε
E hj (X (tj ), α (tj )) ϕ00xx (Xk , αlmε )
j=1 l=t/∆ε k:k∈χεl (41)
)
× µ(Xk , αk )ξ(Xk , αk )ε = 0.
Using the boundedness of ϕ00xx (·, i), hj (·, i), the linear growth condition (15)
for µ(·, i), and the estimate (26), we have,

(t+s)/∆ε −1
κ
Y
hj (X ε (tj ), αε (tj )) ϕ00 2 2
X X
E xx (Xk , αlmε )µ (Xk , αk )ε
j=1 l=t/∆ε k:k∈χε l

(t+s)/∆ε −1 h i (t+s)/∆ε −1
X h i
ε2 E µ2 (Xk , αk ) ≤ Kε2 1 + E Xk2
X X X
≤K
l=t/∆ε k:k∈χεl l=t/∆ε k:k∈χεl
T 2
≤ Kε · = O(ε) → 0 as ε → 0.
ε
(42)
In view of (32), we have
 
Yκ (t+s)/∆ε −1 
ε ε
ϕ00 2
X X
E hj (X (tj ), α (tj )) xx (Xk , αlmε )ξ (Xk , αk )
 
j=1
 l=t/∆ε k:k∈χεl 
Yκ (t+s)/∆ε −1 h i
hj (X ε (tj ), αε (tj )) ϕ00 2
X X
=E xx (Xk , αlmε )Ek ξ (Xk , αk )
 ε

 
Yκ (t+s)/∆ε −1 
hj (X ε (tj ), αε (tj )) ϕ00 2
X X
=E xx (Xk , αlmε )σ (Xk , αk )ε .
 ε

210 R. H. Liu
By using a similar argument to the one leading to (39), we can show that
 
Yκ (t+s)/∆ε −1 
X X
lim E hj (X ε (tj ), αε (tj )) ϕ00 2
xx (Xk , αlmε )ξ (Xk , αk )
ε→0  ε

  (43)
Yκ Z t+s 
2
=E hj (X(tj ), α(tj )) ϕ00
xx (X(u), α(u))σ (X(u), α(u)) du .

j=1 t 
Using (38)–(43) in (37), we obtain

 
Yκ (t+s)/∆ε −1
ε ε
X 
lim E hj (X (tj ), α (tj )) ϕ(X(l+1)mε , αlmε ) − ϕ(Xlmε , αlmε )
ε→0  
j=1 l=t/∆ε
( κ
Y Z t+s h
=E hj (X(tj ), α(tj )) ϕ0x (X(u), α(u))µ(X(u), α(u))
j=1 t
)
i
2
+ ϕ00
xx (X(u), α(u))σ (X(u), α(u)) du .
(44)
Step 3. Now we work on the first summation (the second line) of (36). Using
the continuity of ϕ(·, i) for each i ∈ M, we can argue along the same line
as in Step 2 that the limit of
(t+s)/∆ε −1
X
ϕ(X(l+1)mε , α(l+1)mε ) − ϕ(X(l+1)mε , αlmε )
l=t/∆ε
is the same as that of

(t+s)/∆ε −1
X
ϕ(Xlmε , α(l+1)mε ) − ϕ(Xlmε , αlmε ) + ηε
l=t/∆ε
where ηε → 0 in probability uniformly in t, as ε → 0. We then have,

 
Y κ (t+s)/∆ε −1 
X
hj (X ε (tj ), αε (tj ))

E ϕ(Xlmε , α(l+1)mε ) − ϕ(Xlmε , αlmε )
 
j=1 l=t/∆ε
 
Yκ (t+s)/∆ε −1 
X X
=E hj (X ε (tj ), αε (tj )) [ϕ(Xlmε , αk+1 ) − ϕ(Xlmε , αk )]
 
κ (t+s)/∆ε −1 m0
(
Y X X X
hj (X ε (tj ), αε (tj ))

=E Ek [ϕ(Xlmε , αk+1 )
j=1 l=t/∆ε k:k∈χεl i=1
)

− ϕ(Xlmε , αi )]I{αk =i}
(t+s)/∆ε −1 m0
( κ
"m
Y X X X X 0
=E hj (X ε (tj ), αε (tj )) ϕ(Xlmε , j)

j=1 l=t/∆ε k:k∈χεl i=1 j=1
# )
× P {αk+1 = j|αk = i} − ϕ(Xlmε , i) I{αk =i}
(t+s)/∆ε −1 m0
( κ
"m
Y X X X X 0
ε ε
=E hj (X (tj ), α (tj )) ϕ(Xlmε , j)(δij + εqij )
# )
− ϕ(Xlmε , i) I{αk =i}
 
Yκ (t+s)/∆ε −1 m0 X
m0 
X X X
=E hj (X ε (tj ), αε (tj )) εqij ϕ(Xlmε , j)I{αk =i}
 
κ
(
Y Z t+s
→E hj (X(tj ), α(tj )) Qϕ (X(u), ·) (α(u)) du as ε → 0,
j=1 t
(45)
in which (21) is used for the one-step transition probability of {αn , n ≥ 0}.
Step 4. By combining Step 1 to 3, we have

 
Yκ 
ε ε ε ε ε ε
E hj (X (tj ), α (tj )) [ϕ(X (t + s), α (t + s)) − ϕ(X (t), α (t))]
 
j=1
 
Yκ Z t+s 
→E hj (X(tj ), α(tj )) Lϕ(X(u), α(u)) du as ε → 0.

j=1 t 
212 R. H. Liu
On the other hand, by virtue of the weak convergence of (X ε (·), αε (·)), the
Skorohod representation, and the continuity of hj (·, i) and ϕ(·, i), we have,
 
Yκ 
hj (X ε (tj ), αε (tj )) ϕ(X ε (t + s), αε (t + s)) − ϕ(X ε (t), αε (t))

E
 
j=1
 
Y κ 
→E hj (X(tj ), α(tj )) [ϕ(X(t + s), α(t + s)) − ϕ(X(t), α(t))] as ε → 0.
 
j=1
Hence Equation (34) is verified, and the desired result follows and the proof
of Theorem 1 is completed.
4. Concluding Remarks
In this paper we present a multinomial tree method for option pricing
when the underlying asset price follows a regime-switching model. The new
tree grows linearly as the number of time steps increases. We prove the
weak convergence of the discrete tree approximations to the continuous
regime-switching process by applying the method of martingale problem
formulation.
References
1. D.D. Aingworth, S.R. Das and R. Motwani, A simple approach for pricing
equlty options with Markov switching state variables, Quantitative Finance,
6 (2006), 95-105.
2. C. Albanese, Affine lattice models, International Journal of Theoretical and
Applied Finance, 8 (2005), 223-238.
3. N.P.B. Bollen, Valuing options in regime-switching models, Journal of Deriva-
tives, Vol. 6 (1998), 38-49.
4. P. Boyle and T. Draviam, Pricing exotic options under regime switching,
Insurance: Mathematics and Econimics, Vol. 40 (2007), 267-282.
5. J. Buffington and R.J. Elliott, American options with regime switching, Int.
J. Theor. Appl. Finance, 5 (2002), 497-514.
6. L.B. Chang and K. Palmer, Smooth convergence in the binomial model,
Finance and Stochastics, Vol. 11, No. 1 (2007), 91-105.
7. J. Cox, S. Ross, and M. Rubinstein, Option pricing, a simplified approach,
Journal of Financial Economics 7 (1979), 229-263.
8. F. Diener and M. Diener, Asymptotics of the price oscillations of a european
call option in a tree model, Mathematical Finance, 14 (2004), 271-293.
9. P. Eloe, R.H. Liu and J.Y. Sun, Double barrier option under regime-
switching exponential mean-reverting process, International Journal of Com-
puter Mathematics, Vol. 86, No. 6 (2009), 964-981.
10. I. Florescu and F.G. Viens, Stochastic volatility: Option pricing using a multi-
nomial recombining tree, Applied Mathematical Finance, 15 (2008), 151-181.
11. X. Guo, Information and option pricings, Quant. Finance, Vol. 1 (2000),
38-44.
12. X. Guo and Q. Zhang, Closed-form solutions for perpetual American put
options with regime switching, SIAM J. Appl. Math., 64, No. 6 (2004), 2034-
2049.
13. M.R. Hardy, A regime-switching model for long-term stock returns, North
American Actuarial Journal, Vol. 5, No. 2 (2001), 41-53.
14. H. He, Convergence from discrete to continuous time contingent claims prices,
The Review of Finanical Studies, Vol. 3 (1990), 523-546.
15. S. Heston and G.F. Zhou, On the rate of convergence of discrete-time con-
tingent claims, Mathematical Finance, 10 (2000), 53-75.
16. J.E. Hilliard and A. Schwartz, Binomial option pricing under stochastic
volatility and correlated state variables, Journal of Derivatives, Fall (1996),
23-39
17. A.Q.M. Khaliq and R.H. Liu, New numerical scheme for pricing American
option with regime-switching, Int. J. Theor. Appl. Finance, Vol. 12, No. 3
(2009), 319-340.
18. H.J. Kushner, Approximation and Weak Convergence Methods for Random
Processes with Applications to Stochastic Systems Theory, MIT Press, Cam-
bridge, MA, 1984.
19. R.H. Liu, Regime-switching recombining tree for option pricing, Int. J.
Theor. Appl. Finance, Vol. 13, No. 3 (2010), forthcoming.
20. R.H. Liu, Q. Zhang and G. Yin, Option pricing in a regime switch-
ing model using the Fast Fourier Transform, Journal of Applied
Mathematics and Stochastic Analysis, Vol. 2006, Article ID 18109,
doi:10.1155/JAMSA/2006/18109.
21. D.B. Nelson and K. Ramaswamy, Simple binomial processes as diffusion ap-
proximations in financial models, The Review of Finanical Studies, Vol. 3,
No. 3 (1990), 393-430.
22. J.A. Primbs, M. Rathinam and Y. Yamada, Option pricing with a pen-
tanomial lattice model that incorporates skewness and kurtosis, Applied
Mathematical Finance, 14 (2007), 1-17.
23. J.B. Walsh, The rate of convergence of the binomial tree scheme, Finance
and Stochastics, Vol. 7, No. 3 (2003), 337-361.
24. D.D. Yao, Q. Zhang and X.Y. Zhou, A regime-switching model for European
options, in Stochastic Processes, Optimization, and Control Theory: Appli-
cations in Financial Engineering, Queueing Networks, and Manufacturing
Systems (2006), H.M. Yan, G. Yin, and Q. Zhang Eds., 281-300, Springer.
25. G. Yin, Q. Song, and Z. Zhang, Numerical solutions for jump-diffusions with
regime switching, Stochastics, 77 (2005), 61-79.
26. G. Yin and Q. Zhang, Continuous-Time Markov Chains and Applications: A
Singular Perturbation Approach, Springer, 1998.
214 R. H. Liu
27. G. Yin and Q. Zhang, and G. Badowski, Discrete-time singularly perturbed

Markov chains: Aggregation, occupation measures, and switching diffusion
limit, Advances in Applied Probability, 35 (2003), 449-476.
28. G. Yin and C. Zhu, Hybrid Switching Diffusions: Properties and Applications,
Springer, 2010.
May 11, 2011 13:54 WSPC - Proceedings Trim Size: 9in x 6in 09-luo
215
Optimal Reinsurance under a Jump Diffusion Model
Shangzhen Luo
Department of Mathematics,
University of Northern Iowa,
Cedar Falls, Iowa, 50614-0506, USA
Email: luos@uni.edu
We study an optimal dynamic control problem of an insurance company whose

wealth process is modeled by a jump diffusion. The company can purchase
proportional reinsurance to reduce its risk level and invest all surplus in a risk
free asset. The objective is to maximize the expected value of an exponential
utility of the terminal wealth. We derive explicit optimal value function and
find its associated reinsurance strategy.
Keywords: Stochastic control; HJB equation; proportional reinsurance; expo-

nential utility; jump diffusion.
1. Introduction
During last few decades, various applications of stochastic control theory
in actuarial science have been appearing. See, e.g., Asmussen and Tak-
sar,1 Browne,2 Browne,3 Emanuel et. al,5 Höjgaard and Taksar,7 Höjgaard
and Taksar,8 Liang,9 Luo et. al,10 Schmidli,11 Schmidli,12 Taksar and
Markussen13 etc. Specifically, Browne2 studied a diffusion model with in-
vestment under exponential utility maximization and ruin minimization
criteria; Yang and Zhang15 studied the same problem under more general
jump diffusion settings. Noticing that in many recent actuarial models, e.g.,
Höjgaard and Taksar,7 Höjgaard and Taksar,8 and Luo et. al,10 control of
reinsurance purchase is incorporated as a part of the management of the
insurance company. In this paper, we apply stochastic control theory to
solve an optimization problem on reinsurance. We consider an exponential
utility maximization as the optimization scenario; more exactly the objec-
tive is to seek an optimal proportional reinsurance policy to maximize the
expected utility of the terminal wealth; here we assume that the surplus
of the insurance company is modeled by a perturbed compound Poisson
process, which is a jump diffusion process. Note that this class of risk pro-
216 S. Luo
cesses has been studied extensively in the literature (see, e.g., Dufresne and
Gerber,4 Gerber and Yang,6 and Wang and Wu14 ). A similar study to this
paper has been conducted in Liang,9 where the utility maximization prob-
lem was studied under both diffusion approximation and jump diffusion
models. There investment was not considered. However in this paper (as in
Taksar and Markussen13 ) we assume that there is a risk free asset in the
market, and we then solve the optimization problem in the case with all
surplus invested in the riskless asset. More explicit results are given when
the claim size distribution is exponential or gamma. Economic analysis on
relations between optimal reinsurance strategy and exogenous parameters
is also provided.
We start with a perturbed Cramer-Lundberg risk process in differential
form
dRt = µdt + σdwt − dLt ,

(1)
R0 = x,
where µ is the premium rate, x is the initial surplus level, σ is the volatility
of the perturbing process wt which is a standard Brownian motion inde-
pendent of process Lt , and Lt is a compound Poisson process given by
Nt
X
Lt = Yi , (2)
i=1
here Yi ’s are i.i.d. non-negative claims with common distribution function

F , moment generating function MY and finite expectation α, and {Nt }t≥0
is a Poisson process (with intensity λ) that counts the number of claims. We
suppose that the insurance company can buy proportional reinsurance to
reduce the aggregate loss Lt ; however, the risk of the perturbing noise σwt
cannot be controlled by reinsurance, and then the model is never perfectly
hedged. The retention level of proportional reinsurance ut (> 0) at time t is
determined by the management of the insurance company. We denote the
reinsurance control policy by π = {ut }t≥0 . For a fixed retention level ut at
time t, with proportion of ut , the claims are paid by the insurance company,
and the rest proportion of (1 − ut ) by reinsurance company. In return, the
insurance company diverts its premium at rate of β. In this paper, we call
the reinsurance is expensive if β > λα, and cheap if β ≤ λα; we note here
that the quantity λα is the expected loss per unit time. We also assume
that the surplus is invested in a risk free asset with interest rate r, then the
Optimal Reinsurance under a Jump Diffusion Model 217
dynamics of surplus under policy π become

dRtπ = (µ − β(1 − ut ) + rRtπ )dt + σdwt − ut dLt ,
(3)
R0 = x.
The rest of the paper is organized as follows: in Section 2 we formulate
the optimization problem and give its corresponding HJB equation; in Sec-
tion 3 we provide a general solution to the HJB equation; and we consider
the cases when the claim size follows exponential distribution and gamma
distribution in Section 4.
2. Formulation and HJB Equation

Suppose a complete probability space (Ω, F , P ) endowed with filtration
G = {Gt }t≥0 and the standard Brownian motion {wt }t≥0 is adapted to G.
A policy π = {ut }t≥0 is said to be admissible if {ut }t≥0 is a predictable
process with respect to the filtration G and for each t ≥ 0, 0 ≤ ut ≤ 1. The
set of all admissible policies is denoted by Π.
Now consider an exponential utility function
δ −γx
U (x) = c − e , (4)
γ
where c, γ(> 0), and δ(> 0) are constants. For any admissible policy π ∈ Π,
define a value function as
Vπ (t, x) = E[U (RTπ )|Rtπ = x], (5)
where 0 ≤ t ≤ T and −∞ < x < ∞. The objective is to find the optimal
value function
V (t, x) = sup Vπ (x), (6)
π∈Π
and an optimal control policy π ∗ such that V (t, x) = Vπ∗ (t, x).
For any admissible control policy π, applying Ito’s lemma to a func-
tion g(t, x) of the jump diffusion process governed by (3), we obtain the
following:
g(t, Rtπ ) − g(0, x)
Z t
1
= {gs (s, Rsπ ) + [µ − β(1 − us ) + rRsπ ]gx (s, Rsπ ) + σ 2 gxx (s, Rsπ )}ds
0 2
Z t
+ gx (s, Rsπ )σdws + Σ [g(s, Rsπ ) − g(s, Rs− π
)],
0 s∈JL ;s≤t
(7)
218 S. Luo
where gs , gx , and gxx are partial derivatives, and JL is the set of jump-
ing times of the claim process L := {Ls }s≥0 . Thus the generator of the
controlled process under a fixed policy ut ≡ u is given by
1
Au g(t, x) =gt (t, x) + [µ − β(1 − u) + rx]gx (t, x) + σ 2 gxx (t, x)
2 (8)
+ λE[g(t, x − uY ) − g(t, x)].
Suppose the function V has continuous partial derivatives Vt , Vx and
Vxx , and all limits and expectations can be interchanged, then one can show
that the function V solves an HJB equation given as follows
sup Au V (t, x) = 0, (9)
0≤u≤1
with boundary condition

V (T, x) = U (x). (10)
We shall find a classical solution V with Vx > 0 and Vxx < 0 to the HJB
equation on I where
I = [0, T ] × (−∞, ∞). (11)
3. Solution to the HJB Equation

In this section, we solve the HJB equation. Motivated by Browne3 and Yang
and Zhang,15 we try to fit a solution of the following form
δ
V (t, x) = c − exp{−γxer(T −t) + h(T − t)}, (12)
γ
with boundary condition V (T, x) = u(x) or h(0) = 0, where function h is
to be determined later. Now we compute partial derivatives:
Vt (t, x) =(V (t, x) − c)[−h0 (T − t) + γrxer(T −t) ],
Vx (t, x) =(V (t, x) − c)(−γer(T −t) ), (13)
2 2r(T −t)
Vxx (t, x) =(V (t, x) − c)γ e .
Plug-in the expressions in (13) into HJB equation (9) and simplify, then it
becomes
sup { − h0 (T − t) − γ(µ − β(1 − u))er(T −t)
0≤u≤1
(14)
1
+ γ 2 σ 2 e2r(T −t) + λMY (γer(T −t) u) − λ} = 0.
2
By differentiation, the maximizer over u ∈ (−∞, ∞) in the HJB equation,

denoted by ũ(t) (it does not depend on x), is the unique solution of the
following equation
β
+ E(Y exp{uγer(T −t)Y }) = 0.
− (15)
λ
Thus the maximizer over u ∈ [0, 1] in the HJB equation, denoted by u∗ (t)
is given as follows:

0

 if ũ(t) ≤ 0
∗
u (t) = ũ(t) if 0 < ũ(t) < 1 . (16)

if ũ(t) ≥ 1

1
We note that u∗ (t) and ũ(t) are both continuous and increasing functions
in t.
Theorem 3.1. The optimal value function V in (6) admits form (12), its
associated optimal investment feedback function u∗ (t) is given by (16), and
function h in (12) solves the following equation
h0 (T − t) = − γ(µ − β(1 − u∗ (t)))er(T −t)
1 (17)
+ γ 2 σ 2 e2r(T −t) + λMY (γer(T −t) u∗ (t)) − λ,
2
with boundary condition h(0) = 0.
Lemma 3.1. Under the optimal reinsurance policy π ∗ = {u∗ (t)}, the pro-
∗
cess {V (t, Rtπ )}0≤t≥T is a martingale.
Proof. Apply Ito’s formula to V (s, Rsπ∗ ) and take expectation, then we
have
Z t Z t
π∗ u∗ (s) π∗ ∗
E[V (t, Rt ) − V (0, x)] = A V (s, Rs )ds + E[ σVx (s, Rsπ )dws ],
0 0
= 0,
∗ ∗ ∗
since Au (s)
V (s, Rsπ ) = 0 and Vx (s, Rsπ ) is squarely integrable.
One can also show that {V (t, Rtπ )}0≤t≥T is a super-martingale for any ad-
missible control π, and the theorem follows immediately.
Remark 3.1. In the case of cheap reinsurance, i.e., β ≤ λα, it always holds
that ũ(t) ≤ 0. Thus the optimal reinsurance strategy is
u∗ (t) ≡ 0;
220 S. Luo
it implies that the optimal control policy for the insurance company is
to buy 100% of reinsurance, and hence the risk on claims is eliminated.
Consequently, the optimal value function is given as follows
δ σ 2 γ 2 2r(T −t)
exp{−γxer(T −t) +
V (t, x) = c − e
γ 4r
γ σ2 γ 2 γ
− (µ − β)er(T −t) − + (µ − β)},
r 4r r
which does not depend on the claim size distribution.
In the following section, we shall consider the case of expensive reinsurance,

i.e., β > λα, for two special claim distributions - exponential distribution
and Gamma distribution.
4. Special Claim Distributions

In this section, we study the cases with exponential claims and gamma
claims. Some economic analysis and financial interpretations are provided.
Certain insights on the interplay between the optimal reinsurance strategy
and the exogenous parameters are also given.
4.1. The Case with Exponential Claims

In this subsection, we study the special case when the claim size distribution
is exponential with mean α. Immediately, Equation (15) becomes
β
= E[Y exp{uγer(T −t)Y }]
λ
1 ∞ −y
Z
= ye α exp{uγer(T −t)y}dy
α 0
1
= ,
α[ α1 − uγer(T −t)]2
thus
q
1 λ
α − αβ
ũ(t) = . (18)
γer(T −t)
Note that when the reinsurance is expensive β > λα, it always holds ũ(t) >
0.
In the sequel, we give three theorems according to the following three
possible parameter cases:
q
1 λ
α− αβ
(1) γerT ≥ 1,
q
1 λ
α− αβ
(2) 0 < γ < 1, and
q q
1 λ 1 λ
α− αβ α− αβ
(3) γ ≥1> γerT .
The theorems are readily derived

q
from Theorem 3.1, and we omit their
1 λ
α−
proofs. We note that if γerT ≥ 1, then ũ(t) ≥ 1 and u∗ (t) = 1; further,
αβ
when u∗ (x) ≡ 1, the function h in (12) solves
h0 (T − t) = g1 (T − t),
where
1 λ
g1 (T − t) = −γµer(T −t) + σ 2 γ 2 e2r(T −t) + − λ. (19)
2 α( α1 − γer(T −t) )
Thus we obtain:
q
1 λ
α− αβ
Theorem 4.1. If γerT ≥ 1, then the optimal reinsurance retention level
is
u∗ (t) ≡ 1,
and the optimal value function is given by
δ 1 − αγ σ 2 γ 2 2r(T −t)
V (t, x) = c − ( r(T −t)
)λ/r exp{−γxer(T −t) + e
γ 1 − αγe 4r
γ σ2 γ 2 γ
− µer(T −t) + λ(T − t) − + µ}.
r 4r r
q
1 λ
α− αβ
Remark 4.1. In Theorem 4.1, the assumption γerT ≥ 1 is equivalent
λ 1 rT
to β ≥ α( 1 −γerT )2 and α − γe > 0. Noticing
α
λ λ
β≥ > = λα,
α( α1 rT
− γe ) 2 α( α1 )2
it indicates that the reinsurance rate β is very expensive. As a result, the

insurance company never purchases reinsurance under such an expensive
rate.
222 S. Luo
Notice that
s
r(T −t) 1 λ
MY (γe ũ(t)) = MY ( − )
α αβ
1
= q
α[ α1 − ( α1 − λ
αβ )]
r
β
= ,
λα
thus when u∗ (t) ≡ ũ(t), the function h in (12) solves
h0 (T − t) = g2 (T − t),
where
r
r(T −t) 1 β λβ
g2 (T − t) = −γ(µ − β)e + σ 2 γ 2 e2r(T −t) − + 2 − λ. (20)
2 α α
q
1 λ
α− αβ
We note here that under constraint 0 < γ < 1, it always holds that
0 < ũ(t) < 1. These discussions lead to the following theorem:
q
1 λ
α− αβ
Theorem 4.2. If 0 < γ < 1, then the optimal reinsurance retention
level is
q
1 λ
α − αβ
∗
u (t) = ũ(t) = ,
γer(T −t)
δ σ 2 γ 2 2r(T −t) γ
V (t, x) = c − exp{−γxer(T −t) + e − (µ − β)er(T −t)
γ 4r r
r
2 2
β λβ σ γ γ
+( − 2 + λ)(T − t) − + (µ − β)}.
α α 4r r
q
1 λ
α− αβ
Remark 4.2. Notice that the condition 0 < γ < 1 is equivalent to
the union of the following two cases:
(1) α ≥ γ1 and λα < β (expensive reinsurance and high expectation of

the claim distribution); and
(2) α < γ1 and λα < β < α( 1 λ−γ)2 (expensive reinsurance but not too
α
expensive and low expectation of the claim distribution).
Under these conditions (high claim size expectation, or low claim size ex-
pectation with not too expensive reinsurance), the optimal policy always
involves certain levels of reinsurance purchase, and it is never 0% or 100%.
Now
q
we consider the qlast case of expensive reinsurance on parameters:
1 λ 1 λ
γ ≥ 1 > γerT
; under this constraint, we compute t0 , with
0 ≤ t0 < T , given as the following,
q
1 λ
1 α − αβ
t0 = ln , (21)
r γ
so that
ũ(T − t0 ) = 1.
Hence ũ(t) ≤ 1 and u∗ (t) = ũ(t) when 0 < t ≤ T − t0 ; ũ(t) > 1 and
u∗ (t) = 1 when T − t0 < t ≤ T . Now we conclude that the function h in
(12) solves the following equation:
(
g1 (T − t) if 0 ≤ T − t ≤ t0
h0 (T − t) = ,
g2 (T − t) if t0 < T − t ≤ T
where g1 and g2 are defined by (19) and (20) respectively. One checks
g1 (t0 ) = g2 (t0 ); thus the piece-wisely defined function above is continuous
in t. The discussions above lead to the following theorem:
q q
1 λ 1 λ
Theorem 4.3. If γ ≥ 1 > γerT , then the optimal reinsurance
retention level is
(
1 if 0 ≤ T − t ≤ t0
u∗ (t) = , (22)
ũ(t) if t0 < T − t ≤ T
δ 1 − αγ λ/r r(T −t) σ 2 γ 2 2r(T −t)

V (t, x) = c − ( ) exp{−γxe + e
γ 1 − αγer(T −t) 4r
γ σ2 γ 2 γ
− µer(T −t) + λ(T − t) − + µ},
r 4r r
for 0 ≤ T − t ≤ t0 , and
224 S. Luo
δ σ 2 γ 2 2r(T −t) γ
V (t, x) = c − exp{−γxer(T −t) + e − (µ − β)er(T −t)
γ 4r r
r r
2 2
β λβ σ γ β λβ βγ rt0
+( − 2 + λ)(T − t) − −( −2 )t0 − e
α α 4r α α r
1
λ −γ γ
− ln αq + µ},
r λ r
αβ
for t0 < T − t ≤ T .
Remark 4.3. Under the condition in Theorem 4.3, the retention level
increases to 100% as time evolves; the optimal strategy is to buy less rein-
surance gradually down to 0% near the terminal time.
To this end, we have considered all the possible cases of the parameters.
4.2. The Case with Gamma Claims

In this subsection, we assume that claim size distribution F is Gamma with
parameters a and b, and mean α = a/b. Under this assumption, equation
(15) becomes
β
= E[Y exp{uγer(T −t)Y }]
λ
1 ∞ ba y a exp{−by + uγer(T −t)y}
Z
= dy
α 0 Γ(a)
ba Γ(a + 1)
=
Γ(a)(b − uγer(T −t))(a+1)
Z ∞
(b − uγer(T −t) )(a+1) y (a+1)−1 exp{−(b − uγer(T −t) )y}
× dy
0 Γ(a + 1)
aba
= ,
(b − uγer(T −t))(a+1)
thus
1
λaba a+1

b− β
ũ(t) = . (23)
γer(T −t)
We note that under expensive reinsurance β > λα = λ ab , ũ is always posi-
tive.
In the following, we give three theorems with respect to three possible
conditions on the parameters characterized by:
1
λaba a+1
b− β
(1) γerT ≥ 1,
1
a+1
λaba
b− β
(2) 0 < γ < 1, and
1
a+1 1
λaba λaba a+1
b− β b− β
(3) γ ≥1> γerT .
1
λaba a+1
b− β
Theorem 4.4. If γerT ≥ 1, then the optimal reinsurance retention
level is
u∗ (t) ≡ 1,
and the optimal value function is given by (12) where h solves the following
equation
h0 (T − t) = f1 (T − t),
where
1 b a
f1 (T − t) = −γµer(T −t) + σ 2 γ 2 e2r(T −t) + λ r(T −t)
− λ, (24)
2 b − γe
Remark 4.4. In Theorem 4.4, the assumption on parameters

1
a a+1
b− λab a
β
γerT ≥ 1 is equivalent to β ≥ (b−γeλab
rT )(a+1) and b − γe
rT
> 0.
We observe that
λaba λaba
β≥ > = λα,
(b − γerT )(a+1) b(a+1)
and it implies that the reinsurance is very expensive. Consequently, the
insurance company should never purchase reinsurance.
Now we compute
λaba a+1
1
MY (γer(T −t) ũ(t)) = MY (b − )
β
b a
= a
1 .
λab
a+1
β
1
a+1
λaba
b− β
Further we notice that under the condition 0 < γ < 1, it always
holds that 0 < ũ(t) < 1, and this leads to the following theorem:
226 S. Luo
1
a+1
λaba
b− β
Theorem 4.5. If 0 < γ < 1, then the optimal reinsurance
retention level is
1
λaba a+1

b− β
u∗ (t) = ũ(t) = ,
γer(T −t)
and the optimal value function is given by (12) and function h solves
h0 (T − t) = f2 (T − t),
where
1 λaba a+1
1
f2 (T − t) = − γ(µ − β)er(T −t) + σ 2 γ 2 e2r(T −t) − β(b − )
2 β
ba (25)
+λ a
a − λ,
λab
a+1
β

1
a+1
λaba
b− β
Remark 4.5. The condition in Theorem 4.5, 0 < γ < 1, is
equivalent to the union of the following two cases:
(1) α ≥ γa and λα < β (expensive reinsurance and high claim size

expectation); and
λaba
(2) α < γa and λα < β < α(b−γ) a+1 (expensive reinsurance but not too
expensive and low claim size expectation).
Under these conditions, the optimal policy always involves certain positive
level of reinsurance purchase that is never 0% or 100%.
Now we consider the last case of expensive reinsurance on parameters:

1
a+1 1
a a a+1
b− λab
β b− λab
β
γ ≥ 1 > γerT ; under this constraint, there exists t1 ,
0 ≤ t1 < T , given as the following,
1
λaba a+1
1 b− β
t1 = ln , (26)
r γ
such that
ũ(T − t1 ) = 1.
Consequently, ũ(t) ≤ 1 and u∗ (t) = ũ(t) when 0 < t ≤ T − t1 ; ũ(t) > 1 and
u∗ (t) = 1 when T − t1 < t ≤ T . Now we come to the following theorem:
1
a+1 1
λaba λaba a+1
b− β b− β
Theorem 4.6. If γ ≥1> γerT , then the optimal rein-
surance retention level is
(
1 if 0 ≤ T − t ≤ t1
u∗ (t) = , (27)
ũ(t) if t1 < T − t ≤ T
and the optimal value function is given by (12), where the function h solves
the following equation:
(
0 f1 (T − t) if 0 ≤ T − t ≤ t1
h (T − t) = ,
f2 (T − t) if t1 < T − t ≤ T
here f1 and f2 are defined by (24) and (25) respectively, and the function
above is continuous.
Under the paramerter condition in Theorem 4.6, the optimal strategy sug-
gests to buy less reinsurance gradually down to 0% to the final time.
The solution of the case with gamma claims is now complete.
4.3. Numerical Examples

In this subsection, we give two examples with plots to demonstrate the
results we obtained in the previous subsections.
Example 4.1. In this example we assume that the claim distribution is

exponential with mean α = 0.4, and the other model parameters are given
as the following: β = 1, λ = 2, γ = 0.2, r = 0.15, σ = 0.5, µ = 0.2, c = 1,
and δ = 0.15. Notice that the reinsurance premium rate β is higher than the
expected loss per unit time λα, and the reinsurance is expensive. Further
assume
q
the surplus level is x = q
1 and expiration time is T = 4; and compute
1 λ 1 λ
γ ≈ 1.3197 ≥ 1 and γerT ≈ 0.7242 < 1; one can apply Theorem
4.3 to obtain the optimal risk exposure (reinsurance intention level) u∗ (t)
and the optimal value function V (t, x). There is a unique threshold point
in time T − t0 ≈ 2.1508 at which the form of the optimal risk exposure
function changes. The plot is given in Figure 1, where u∗ (t) is an increasing
function in t; this means that as time approaches to the expiration date,
less reinsurance is bought (down to 0% at the end).
Example 4.2. In this example we assume a Gamma claim distribution,

whose parameters are a = 2 and b = 3. Note that its mean is α = ab = 2/3.
The other parameters are given as below: β = 1, λ = 1, γ = 0.3, r = 0.15,
228 S. Luo
σ = 0.5, µ = 0.2, c = 1, and δ = 0.15. Further assume T = 4 and x = 1.

1
a+1 1
a+1
a a
b− λab
β
b− λab
β
Now compute γ ≈ 1.2642 and γerT ≈ 0.6938. Thus we
can apply Theorem 4.6 to obtain the optimal risk exposure function and the
optimal value function. The plot is drawn in Figure 2. The unique threshold
point is at T − t1 ≈ 2.4371.
1.1
Optimal Value Function
0.9
0.8
Optimal Risk Exposure

0.7
0.6
Value Function Without Reinsurance
0.5
0.4
T−t0
0 0.5 1 1.5 2 2.5 3 3.5 4

Time
Fig. 1. The case with exponential claims (Theorem 4.3)
Acknowledgments
This research is supported by a UNI summer research fellowship.
References
1. Asmussen, S. and Taksar, M.: Controlled Diffusion Models for Optimal Div-
idend Payout, Insurance: Math. Econ. 20, 1–15 (1997)
1.1
1.05
0.95
0.9
0.85
0.8
Optimal Value Function
0.75
Optimal Risk Exposure
0.7
0.65
T−t1
0 0.5 1 1.5 2 2.5 3 3.5 4

Time
Fig. 2. The case with Gamma claims (Theorem 4.6)
2. Browne, S.: Optimal Investment Policies for a Firm with a Random Risk
Process: Exponential Utility and Minimizing the Probaiblity of Ruin, Math.
Ope. Res 20(4), 937–958 (1995)
3. Browne, S.: Survival and Growth with Liability: Optimal Portfolio Strategies
in Continuous Time, Math. Ope. Res 22, 468–492 (1997)
4. Dufresne, F. and Gerber, H. U. Risk theory for a compound Poisson process
that is perturbed by diffusion. Insurance Math. Econom. 10 51–59 (1991)
5. Emanuel, D.C., Harrison J.M. and Taylor, A.J.: A Diffusion Approximation
for the Ruin Probability with Compounding Assets, Scan. Actuarial J. 37–45
(1975)
6. Gerber, H.U. and Yang, H.L.: Absolute Ruin Probabilities in a Jump Diffu-
sion Risk Model with Investment, North American Actuarial Journal 11(3),
159–169 (2007)
7. Höjgaard., B. and Taksar, M.: Optimal Proportional Reinsurance Policies
for Diffusion Models with Transaction Costs, Insurance Math. Econom. 22,
41–51 (1998)
8. Höjgaard., B. and Taksar, M.: Optimal Proportional Reinsurance Policies for
Diffusion Models, Scan. Actuarial J. 166–180 (1998)
230 S. Luo
9. Liang, Z.: Optimal Proportional Reinsurance for Controlled Risk Process

which is Perturbed by Diffusion, Acta Mathematicae Applicatae Sinica 23(3),
477-488 (2007)
10. Luo, S., Taksar, M. and Tsoi, A.H.: On Reinsurance and Investment for Large
Insurance Portfolios, Insurance Math. Econom. 42 (1) 434–444 (2008)
11. Schmidli, H.: Optimal Proportional Reinsurance Policies in a Dynamic Set-
ting, Scan. Actuarial J. 55–68 (2001)
12. Schmidli, H.: On Minimizing The Ruin Probability by Investment and Rein-
surance, The Annals of Applied Probability 12(3), 890–907 (2002)
13. Taksar, M. and Markussen C.: Optimal Dynamic Reinsurance Policies for
Large Insurance Portfolios, Finance and Stochastics 7 97–121 (2003)
14. Wang, G.J. and Wu, R.: Some distributions for classical risk process that is
perturbed by diffusion, Insurance Math. and Econom. 26(1) 15–24 (2000)
15. Yang, H.L. and Zhang, L.H.: Optimal Investment for Insurer with Jump-
diffusion Risk Process Insurance Math. Econom. 37 615–634 (2005)
April 26, 2011 17:27 WSPC - Proceedings Trim Size: 9in x 6in 10-sun
231
Applications of Counting Processes and Martingales in Survival

Analysis
Jianguo Sun
Department of Statistics
University of Missouri, USA
Email: sunj@missouri.edu
The counting process and martingale have been the research topic in probabil-
ity for a long time and there exists a great deal of literature for both their stud-
ies and their applications. This article discusses their applications in survival
analysis, a field that plays an essential and fundamental role in biostatistics
and deals with experiments concerning times to events. In addition to being a
key component in most of medical and public health studies, survival analysis
is applied to many other fields including demographical studies, economics, re-
liability studies and social studies. Although survival analysis has existed for a
long time, the modern survival analysis really started about 30 years ago when
the counting process and martingale tools were applied for the advance of the
field (Aalen, 1975, 1980; Andersen et al., 1993).
Keywords: Counting processes; Cox model; martingales; partial likelihood;

right-censored failure time data; semiparametric regression analysis.
1. Introduction
Survival analysis refers to the statistical field that deals with data con-
cerning positive random variables representing times to certain events
(Kalbfleisch and Prentice, 2002). Examples of the event, often referred to
as the failure or survival event, include death, the onset of a disease or
certain milestone, the failure of a mechanical component of a machine, or
learning something. The occurrence of the event is usually referred to as a
failure and often we also use the terminology failure time data analysis and
refer to the variable of interest as failure or survival time or variable. Fail-
ure time or survival data arise extensively in medical studies, but there are
many other investigations that also produce survival data. These include
biological studies, demographical studies, economic and financial studies,
epidemiological studies, psychological experiments, reliability experiments,
and sociological studies.
232 J. Sun
Table 1. Remission times in weeks for acute leukemia patients
Group Survival times in weeks
6-MP 6, 6, 6, 6∗ , 7, 9∗ , 10, 10∗ , 11∗ , 13, 16, 17∗ , 19∗ , 20∗ , 22, 23, 25∗
32∗ , 32∗ , 34∗ , 35∗
Placebo 1, 1, 2, 2, 3, 4, 4, 5, 5, 8, 8, 8, 8, 11, 11, 12, 12, 15, 17, 22, 23
Survival analysis differs from other statistical fields due to many rea-
sons. One major reason, also the feature that distinguishes the analysis of
failure time data from other statistical problems, is the existence of cen-
soring such as right censoring that will be introduced below. Censoring
mechanisms can be quite complicated and thus necessitate special methods
of treatment. The methods available for other types of data are usually
simply not appropriate for censored data. Truncation is another feature of
some failure time data that requires special treatments. Because of censor-
ing and/or truncation, failure time data are always incomplete and the goal
of survival analysis is to develop statistical approaches that can handle such
incompleteness.
To give an example, Table 1 presents a typical set of failure time data,
which arose from a clinical trial on acute leukemia patients. The table
gives remission times in weeks for 42 patients in two treatment groups.
One treatment is the drug 6-mercaptopurine (6-MP) and the other is the
placebo treatment. The study was performed over a one-year period and
the patients were enrolled into the study at different times. In the table, the
starred numbers are right-censored remission times, meaning that the true
remission times were known only to be greater than these censored times.
This happened because these patients were still in the state of remission
at the end of the trial and the times observed are the amount of time
from when the patient entered the study to the end of the study. For other
patients, their remission times were observed exactly.
In survival analysis, it is common to use T to denote the nonnegative
random variable under study or the failure time of interest. For inference
about T , unlike other statistical fields, it is more common and convenient
to consider or model the so-called survival and hazard functions defined as
S(t) = P ( T > t ) , 0 < t < ∞

and
P ( t ≤ T < t + ∆t | T ≥ t )
λ(t) = lim ,
∆t→0+ ∆t
May 13, 2011 11:28 WSPC - Proceedings Trim Size: 9in x 6in 10-sun
Applications of Counting Processes and Martingales 233
respectively, assuming that T is a continuous variable. Here we will focus on

continuous survival variables and the hazard function is defined differently
for discrete survival variables (Kalbfleisch and Prentice, 2002). It can be
easily shown that
Z t
f (t) d log S(t)
λ(t) = =− , S(t) = exp − λ(s) ds = exp [ − Λ(t) ] ,
S(t) dt 0
Rt
where f (t) denotes the density function of T and Λ(t) = 0 λ(s) ds, usually
referred to as the cumulative hazard function.
A counting process, usually denoted by N (t), is a stochastic process
that can be thought of as registering the occurrences in time of certain
recurrent events. That is, N (t) is a step function jumping by one each
time when the event occurs. In survival analysis, more often one will face a
multivariate counting process, a vector of counting processes all being zero
at time zero with paths which are piecewise constant and nondecreasing,
having jumps of size one only with no two processes jumping simultaneously.
A martingale is also a stochastic process often regarded as a kind of pure
random noise process. Many stochastic processes can be written as the sum
of a martingale and a finite variation predictable stochastic process with
the latter called the compensator of the process. In the following, we will
show that many inference problems in survival analysis can be conveniently
formulated using the counting process and martingale and thus one can
apply the existing, rich theory about them to solve the problems much
easily.
The remainder of the paper is organized as follows. In Section 2, we will
consider regression analysis of right-censored failure time data using the
Cox or proportional hazards model. Section 3 discusses regression analysis
of current status failure time data using the additive hazards model. Current
status failure time data mean that the failure time of interest is observed
only to be either smaller or larger than a number, usually the observation
time. In this case, each study subject is observed only once and one area that
often produces such data is cross-sectional studies. In Section 4, regression
analysis of bivariate current status failure time data will be described under
the proportional hazards model. For all situations, the focus will be on
estimation of regression parameters. Section 5 contains some concluding
remarks.
234 J. Sun
2. Regression Analysis of Right-censored Failure Time

Data Using the Cox Model
Consider a survival study that consists of n independent subjects. For sub-
ject i, let Ti denote the failure time of interest and Zi a vector of covariates.
In reality, it is usually the case that there exists a so-called censoring vari-
able Ci such that one may only observe Ci and Ti > Ci . This is usually
referred to as right-censoring and one common situation where this hap-
pens is that Ci represents the end of the study and by then, the subject is
still alive or the event of interest has not occurred yet. In other words, the
observed data have the form
{ (Xi = min(Ti , Ci ), δi = I(Xi = Ti ), Zi ; i = 1, ..., n } .
Suppose that the goal of the study is to assess the effect of covariates on
Ti or estimate the covariate effect. For this, a common approach is to specify
a regression model and in survival analysis, perhaps the most commonly
used regression model is the Cox or proportional hazards model defined as
λ(t|Zi ) = λ0 (t) exp( Zi0 β) (1)
(Cox, 1972). Here λ(t|Zi ) denotes the hazard function of Ti given Zi , λ0 (t) is
an unknown baseline hazard function, and β denotes the vector of regression
parameters.
To estimate or make inference about β, the common approach is to
maximize the partial likelihood
n
" #δi
Y exp(Zi0 β)
L(β) = P 0
i=1 Xj ≥Xi exp(Zj β)
or to solve the partial score function U (β) = ∂ log L(β)/∂β = 0. This ap-
proach was originally proposed by Cox (1975) and has been studied by many
authors (Andersen et al., 1993; Kalbfleisch and Prentice, 2002) using the
formulation and theory of the counting process and martingale described
below. A key advantage of this partial likelihood approach is that L(β) is
independent of the baseline hazard function λ0 (t). In other words, one does
not have to deal with or estimate λ0 (t).
To express L(β) or U (β) using the counting process and martingale,
define
Nri (t) = I(Xi ≤ t, δi = 1)
and Yri (t) = I(Xi ≥ t), i = 1, ..., n. It is easy to see that Nri (t) is a counting
process jumping from 0 to 1 at t = Ti if the survival time Ti is observed
instead of the censoring time Ci . Yri (t) is usually referred to as the indicator
process indicating if subject i is under study at time t. Also define
Z t
Mri (t) = Nri (t) − Yri (u) exp(Zi0 β)λ0 (u)du , i = 1, ..., n ,
0
n
Zij Yri (t) exp(Zj0 β) , j = 0, 1,
X
Sr(j) (β, t) =n −1
i=1
(1) (0)
and Er (β, t) = Sr (β, t)/Sr (β, t). Then {Mri (t)} are martingales and the
partial likelihood function can be rewritten as L(β) = L(β, ∞) with
n
" #∆Nri (u)
Y (u) exp(Zi0 β)
Pn ri
Y Y
L(β, t) = 0 .
i=1 0≤u≤t j=1 Yrj (u) exp(Zj β)
Furthermore, we have U (β) = Ur (β, ∞) with

n t
∂ log L(β, t) X
Z
Ur (β, t) = = [Zi − Er (β, u)] dNri (u)
∂β i=1 0
n Z
X t
= [Zi − Er (β, u)] dMri (u) . (2)
i=1 0
Let β̂r denote the solution to U (β, ∞) = 0. It is clear that the integrand
with respect to each Mri (t) is a predictable process and thus Ur (β, t) is the
sum of martingales. By applying the martingale theory and using expression
(2), one can show that β̂r is a consistent estimate of β in the proportional
hazards model (1). Furthermore, Ur (β, t) converges to a Gaussian process
and β̂r converges in distribution to a normal random vector. The details
can be found in Andersen and Gill (1982).
3. Regression Analysis of Current Status Failure Time

Data Using the Additive Hazards Model
In this section, instead of right censoring, we will discuss failure time data
with a different type of censoring. First we will introduce current status
failure time data and then consider their regression analysis.
236 J. Sun
Table 2. Death times in days for 144 male RFM mice with lung tumors
Group Tumor status Death times
CE With tumor 381, 477, 485, 515, 539, 563, 565, 582, 603, 616, 624, 650
651, 656, 659, 672, 679, 698, 702, 709, 723, 731, 775, 779
795, 811, 839
No tumor 45, 198, 215, 217, 257, 262, 266, 371, 431, 447, 454, 459
475, 479, 484, 500, 502, 503, 505, 508, 516, 531, 541, 553
556, 570, 572, 575, 577, 585, 588, 594, 600, 601, 608, 614
616, 632, 632, 638, 642, 642, 642, 644, 644, 647, 647, 653
659, 660, 662, 663, 667, 667, 673, 673, 677, 689, 693, 718
720, 721, 728, 760, 762, 773, 777, 815, 886
GE With tumor 546, 609, 692, 692, 710, 752, 773, 781, 782, 789, 808, 810
814, 842, 846, 851, 871, 873, 876, 888, 888, 890, 894, 896
911, 913, 914, 914, 916, 921, 921, 926, 936, 945, 1008
No tumor 412, 524, 647, 648, 695, 785, 814, 817, 851, 880, 913, 942
986
3.1. Current Status Failure Time Data

As mentioned above, current status failure time data refer to the failure time
data in which each failure time is known or observed to be either smaller
or larger than a number, usually the observation time. The term current
status data originates from demographical studies and such data are also
sometimes referred to case I interval-censored data (Sun, 2006). Current
status data occur when each study subject is observed only once and the
only observed information for the survival event of interest is whether the
event has occurred before or after the observation time, which in survival
analysis is also referred to as either left- or right-censored. In addition to
cross-sectional studies, tumorigenicity experiments are another field where
one often faces current status failure time data.
Table 2 gives a standard set of current status data arising from a tu-
morigenicity experiment of male RFM mice concerning lung tumors (Hoel
and Walberg, 1972). In the experiment, the animals were put in one of two
environments, conventional environment (CE) and germ-free environment
(GE), and one of the goals was to assess the environment effects on the
growth rate of lung tumors. In other words, the failure time of interest
is the time to tumor onset. Unfortunately theses times or the occurrences
of the tumors cannot be observed while they were alive. The information
about these times that can observed is the death or sacrifice time and if
lung tumor was present or absent at the death or sacrifice. The sacrifice
here means that the animals were killed, which is often performed to ob-
tain better information about the tumor growth.
3.2. Regression Analysis using the Additive Hazards Model

Consider a survival study that consists of n independent subjects and in
which each study subject is observed or examined only at one time point
for the survival event of interest. For subject i, let Ti and Zi be defined as
before and Ci denote the observation time. Then the observed data have
the form
{ Ci , δi = I(Ti ≤ Ci ), Zi ; i = 1, ..., n } .
For regression analysis, as mentioned before, the proportional hazards

model is the most commonly used but it may not fit data well sometimes.
Also depending on the nature of the study, different models may be pre-
ferred that describe different aspects rather than those specified by the
proportional hazards model. One such model is the additive hazards model
defined as
λ(t|Zi ) = λ0 (t) + Zi0 β (3)
(Lin and Ying, 1994, 1995), where λ0 (t) and β are defined as in model (1).
It can be easily seen that in model (1), β represents the multiplicative effect
of covariates on the hazard function, while β in model (3) gives the additive
effect of covariates, which are often more interesting in fields such as social
studies.
To estimate β in model (3), define Yci (t) = I(Ci ≤ t) and the counting
process
Nci (t) = I{Ci ≤ min(Ti , t)}
for subject i, i = 1, ..., n. Then one can show that Nci (t) has the intensity
Yci (t) exp(−tZi0 β)λc (t)e−Λ0 (t) ,
where λc (t) denotes the hazard function of Ci . Furthermore, the martingale

corresponding to Nci (t) can be defined as
Z t
Mci (t) = Nci (t) − Yci (u) exp(−uZi0 β)λc (u)e−Λ0 (u) du .
0
Now we can develop the estimating equation for estimation of β using

Nci (t) and Mci (t). For this, define
n
X
Sc(j) (β, t) = n−1 Yci (t)(tZi )j exp(−tZi0 β) j = 0, 1 .
i=1
238 J. Sun
Then motivated by Ur (β, t) given in Section 2, one can construct the fol-
lowing estimating function
n Z t
" #
(1)
X Sc (β, u)
Uc (β, t) = tZi − (0) dNci (u)
i=1 0 Sc (β, u)
n Z
" #
t (1)
X Sc (β, u)
= tZi − (0)
dMci (u)
i=1 0 Sc (β, u)
and estimate β by the solution to Uc (β, ∞) = 0. Let β̂c denote the such
defined estimate. Then as β̂r , by applying the counting and martingale
theory, one can prove that β̂c is consistent and converges in distribution to
a normal random vector. The detailed derivation and proofs of the results
given above can be found in Lin et al. (1998).
4. Regression Analysis of Bivariate Current Status Failure

Time Data Using the Cox Model
4.1. Assumptions and the Likelihood Function
In many survival studies, there may exist two related failure time variables
of interest instead of only one variable and in this case, it is obvious that one
has to conduct bivariate or joint analysis of the variables or the observed
data about them. Let T1 and T2 denote two failure time variables of interest
on a study subject and Z the associated vector of covariates. Suppose that
during the study, each study subject is examined or observed only once at
time C. That is, only current status data are available for both T1 and T2 .
Note that in general, there may exist different observation times for T1 and
T2 , but for the simplicity, here we assume that the observation times for
them are the same. Define
δ1 = I(T1 ≤ C) , δ2 = I(T2 ≤ C) ,
indicating if the survival events represented by T1 and T2 have occurred

before C. Then the observed data have the form
(C, δ1 , δ2 , Z)
for one subject.

For the analysis, instead of the partial likelihood and estimating equa-
tion approaches applied in Sections 2 and 3, respectively, we will consider
the full likelihood approach here. For this, we will assume that T1 and T2
follow the copula model specified by
S(s, t) = P (T1 > s, T2 > t) = Cα (S1 (s), S2 (t)) (4)
for the joint survival function of T1 and T2 , where Cα is a distribution

function on the unit square, α ∈ R is a global association parameter, and
S1 and S2 denote the marginal survival functions of T1 and T2 , respectively.
The copula model has been widely used in modeling bivariate variables in
probability and statistics (Hougaard, 2000) and one attractive feature of
model (4) is its flexibility in that it includes as special cases many useful
bivariate failure time models. For example, it gives the Archimedean copula
family by taking
Cα (u, v) = φα {φ−1 −1
α (u) + φα (v)} , 0 ≤ u , v ≤ 1 ,
where 0 ≤ φα ≤ 1, φα (0) = 1, φ0α < 0, φ00α > 0. More specifically, if

letting φα (u) = (1 + u)1/(1−α) , the Laplace transformation of a gamma
distribution, one has
Cα (u, v) = (u1−α + v 1−α − 1)1/(1−α) , α > 1 ,
which is commonly referred to as the Clayton family (Clayton, 1978). An-

other attractive feature of copula models is that the marginal distributions
do not depend on the choice of the association structure. Thus one can
model the marginal distributions and the association separately.
To derive the full likelihood function, define the following counting pro-
cesses
N11 (t) = δ1 δ2 I(C ≤ t) , N10 (t) = δ1 (1 − δ2 ) I(C ≤ t) ,
N01 (t) = (1 − δ1 ) δ2 I(C ≤ t) , N00 (t) = (1 − δ1 ) (1 − δ2 ) I(C ≤ t) .
Then the log likelihood function is proportional to

1 X
X 1 Z ∞
l(α, S1 , S2 ) = log Sjm (θ, t) dNjm (t)
j=0 m=0 0
based on one subject, where
S11 (θ, t) = P (T1 ≤ t, T2 ≤ t) , S01 (θ, t) = P (T1 > t, T2 ≤ t) ,
S10 (θ, t) = P (T1 ≤ t, T2 > t), S00 (θ, t) = P (T1 > t, T2 > t) .
240 J. Sun
4.2. Efficient Estimation of Regression Parameters

For inference about covariate effects, we need to specify a regression model
and for this, one way is to assume that both T1 and T2 follow the propor-
tional hazards model (1) with the same regression parameter β and different
baseline hazard functions λ1 (t) and λ2 (t), respectively. It is straightforward
to generalize the idea and method discussed here to the situation where the
regression parameters for T1 and T2 are different. Under the proportional
hazards model, we have
S1 (t) = exp(−Λ1 (t) exp(Z 0 β)) , S2 (t) = exp(−Λ2 (t) exp(Z 0 β)) ,
Rt Rt
where Λ1 (t) = 0 λ1 (s)ds and Λ2 (t) = 0 λ2 (s)ds are the cumulative base-
line hazard functions of T1 and T2 , respectively. The log likelihood function
l(α, S1 , S2 ) then becomes a function of θ = (β 0 , α)0 and Λ1 (t) and Λ2 (t).
For estimation of θ based on the log likelihood function l(θ, Λ1 , Λ2 ), a
common approach is to derive the efficient score function for θ. For this, let
Dα , Du and Dv denote the derivatives of Cα (u, v) with respect to α, u and
v and at u = S1 and v = S2 , respectively. For j, m = 0, 1, define
(1)
ajm = −(−1)j exp(Z 0 β) Λ1 S1 [m + (−1)m Du ] ,
(2)
ajm = −(−1)m exp(Z 0 β) Λ2 S2 [j + (−1)j Dv ] ,
(1) (2)
and Bjm = (ajm + ajm )−1 (−1)m+j Dα . Also for j, m = 0, 1, let ajm =
(1) (2)
(ajm , ajm )0 and define the matrix

Z Z
Zjm =
Bjm Bjm
and b∗ = H ∗ G−1 , where,
1 X
X 1
E a⊗2

G = jm Y λc /Sjm
j=0 m=0
and
1 X
X 1
E Zjm a⊗2

H = jm Y λc /Sjm .
j=0 m=0
Then one can show that the efficient score function for θ has the form
1 X 1 Z ∞
˙lθ∗ =
X (Zjm − b∗ ) ajm
dNjm (t) (5)
j=0 m=0 0
Sjm
based on a single study subject.

Suppose that the study involves n independent subjects and let Njmi ,
Zjmi , ajmi and Sjmi be defined as Njm , Zjm , ajm and Sjm with respect to
subject i. Then for estimation of β or θ, it is natural to apply the empirical
version of the efficient score function given in (5), which has the form
n X
1 X
1 Z ∞
X (Zjmi − b∗n ) ajmi
l˙θ∗ , n (θ, Λ1 , Λ2 ) = dNjmi (t) .
i=1 j=0 m=0 0 Sjmi
That is, one can estimate θ by the solution to
l˙θ∗ , n (θ, Λ̂1 , Λ̂2 ) = 0 ,
where Λ̂1 and Λ̂2 denote some consistent estimates of Λ1 (t) and Λ2 (t). As
in Sections 2 and 3, one can apply the counting process and martingale
theory to establish the asymptotic properties of the such defined estimate.
Wang et al. (2008) provided more details about the discussion above.
5. Concluding Remarks
In the preceding sections, we discussed regression analysis of failure time
data under three different censoring mechanisms and presented the counting
process and martingale connection to these problems. As it can be seen, for
all three situations, the estimation problems can be expressed using count-
ing processes and martingales and thus one could conveniently employ the
existing theory about them, especially various types of central limit results,
to investigate or establish the asymptotic properties of the resulting esti-
mates commonly expected and needed for inference. It is worth noting that
without the use of the connection to the counting process and martingale
and their theory, many of the asymptotic results could still be obtained
by different approaches but would be much more difficult and complicated.
In contrast, the application of the counting process and martingale makes
these results much easy and straightforward and more importantly, some
of the other results possible and available.
There are many other types of problems in survival analysis in addition
to regression analysis or the assessment of covariate effects and most of
these can be formulated and investigated using the counting process and
martingale. For example, consider the right-censored failure time data dis-
cussed in Section 2 and suppose that one is interested in estimation of the
overall survival function S(t), which is usually the first thing that one does
in analyzing failure time data. Suppose that there do not exist covariates
242 J. Sun
and let Nri and Yri be defined as in Section 2. Define

X n n
X
Nr (t) = Nri (t) , Yr (t) = Yri (t) .
i=1 i=1
It is apparent that Nr (t) is still a counting process. For estimation of S(t),
the most commonly used approach is to use the Kaplan-Meier estimator
(Kaplan and Meier, 1958) that can be expressed as
Y ∆N (s)

Ŝ(t) = 1− .
Y (s)
s≤t
As with estimation of regression parameters, one can easily apply the count-
ing process and martingale theory to investigate the statistical properties
of Ŝ(t) such as deriving the variance estimate. If there is no right censoring
or all δi = 1, then the Kaplan-Meier estimator reduces to one minus the
empirical distribution function.
As discussed above, the Cox or proportional hazards model (1) is the
most commonly used regression model in survival analysis and the additive
hazards model (3) is also a popular choice. There exist many other models
in survival analysis and the selection of an appropriate model for a given
problem usually depends on the nature and prior knowledge of the problem
but could be very difficult sometimes. Note that in both models (1) and
(3), we assumed that covariates are time-independent and actually one can
apply them even if the covariates are time-dependent. Furthermore, both
models assume the linear effects of covariates and one can easily generalize
them to include nonlinear covariate effects. In addition to these two models,
another popular semiparametric model, motivated by the common linear
regression model, is the accelerated failure time model given by
log T = Z 0 β +
(Jin et al., 2003), where denotes a random error with unknown distribution
function. Note that this model directly models the effects of covariates on
the failure time of interest instead of the hazard function as in models
(1) and (3). Sometimes one may be more interested in the direct effects
of covariates on survival function and in this case, one could apply the
proportional odds model given by
SZ (t) S0 (t)
= exp(Z 0 β)
1 − SZ (t) 1 − S0 (t)
(Yang and Prentice, 1999), where SZ (t) denotes the survival function for
a subject with covariates Z and S0 (t) is an unknown baseline survival
function.
In addition to right censoring and current status censoring, there ex-

ist other types of censorings or mechanisms that cause incompleteness in
survival analysis. For example, instead of being observed only once as with
current status censoring, a study subject may be examined or observed a
few times during the study period, which is often the situation in medical
or public health follow-up studies. In these situations, the failure time of
interest could be only observed to belong to a window given by the last
observation time at which the survival event has not occurred and the first
observation time at which the survival event has occurred. This is commonly
referred to as interval censoring (Sun, 2006). Truncation is another feature
that often exists in failure time data and it also causes incompleteness and
needs special treatments as censoring although the mechanism behind it is
usually different from that behind censoring. For example, left-truncation
means that the survival event could be observed or known only if it occurs
after certain time or another event. More discussion on this and all types
of censoring can be found in, for example, Klein and Moeschberger (2003)
and Lawless (2003).
References
Aalen, O. O. (1975). Statistical inference for a family of counting processes.
Ph.D Thesis, University of California, Berkeley.
Aalen, O. O. (1980). A model for nonparametric regression analysis of
counting processes. Spring Lect. Notes Statist. 2, 1-25. Mathemat-
ical Statistics and Probability Theory. W. Klonecki, A. Kozek, and
J. Rosiński, editors.
Andersen, P. K., Borgan, O., Gill, R. D. and Keiding, N. (1993). Statistical
models based on counting processes. Springer-Verlag: New York.
Andersen, P. K. and Gill, R. D. (1982). Cox regression model for count-
ing processes: a large sample study. The Annals of Statistics, 10,
1100-1120.
Clayton, D. G. (1978). A model for association in bivariate life tables and
its application in epidemiological studies of familial tendency in
chronic disease incidence. Biometrika, 65, 141-151.
Cox, D. R. (1972). Regression analysis and life tables (with discussion). J.
R. Statist. Soc. B, 34, 187-220.
Cox, D. R. (1975). Partial likelihood. Biometrika, 62, 269-276.
Hoel, D. G. and Walburg, H. E. (1972). Statistical analysis of survival
experiments. Journal of National Cancer Institute, 49, 361-372.
Hougaard, P. (2000). Analysis of multivariate survival data. Springer-
Verlag: New York.
244 J. Sun
Jin, Z., Lin, D. Y., Wei, L. J. and Ying, Z. (2003). Rank-based inference
for the accelerated failure time model. Biometrika, 90, 341-353.
Kalbfleisch, J. D., and Prentice, R. L. (2002). The statistical analysis of
failure time data. Second edition. Wiley: New York
Kaplan, E. L. and Meier, P. (1958). Nonparametric estimation from incom-
plete observations. Journal of the American Statistical Association,
53, 457-481.
Klein, J. P. and Moeschberger, M. L. (2003). Survival analysis, Springer-
Verlag: New York.
Lawless, J. F. (2003). Statistical models and methods for lifetime data.
John Wiley: New York.
Lin, D. Y., Oakes, D. and Ying, Z. (1998). Additive hazards regression
with current status data. Biometrika, 85, 289-298.
Lin, D. Y. and Ying, Z. (1994). Semiparametric analysis of the additive
risk model. Biometrika, 81, 61-71.
Lin, D. Y. and Ying, Z. (1995). Semiparametric analysis of general
additive-multiplicative hazard models for counting processes. The
Annals of Statistics, 23, 1712-1734.
Sun, J. (2006). The statistical analysis of interval-censored failure time
data. Springer.
Wang, L., Sun, J. and Tong, X. (2008). Efficient estimation for the propor-
tional hazards model with bivariate current status data. Lifetime
Data Analysis, accepted.
Yang, S. and Prentice, R. L. (1999). Semiparametric inference in the pro-
portional odds regression model. Journal of the American Statisti-
cal Association, 94, 125-136.
April 27, 2011 14:8 WSPC - Proceedings Trim Size: 9in x 6in 11-zhang
245
Stochastic Algorithms and Numerics for Mean-Reverting Asset

Trading
Q. Zhang
University of Georgia
Athens, GA 30602, USA
Email: qingz@math.uga.edu
C. Zhuang
Marshall School of Business
University of Southern California
Los Angeles, CA 90089, USA
Email: czhuang@usc.edu
G. Yin
Wayne State University
Detroit, MI 48202, USA
Email: gyin@math.wayne.edu
This work considers trading a mean-reverting asset. The strategy is to deter-

mine a low price to buy and a high price to sell so that the expected return is
maximized. Slippage cost is imposed on each transaction. Our effort is devoted
to developing a recursive stochastic approximation type algorithm to estimate
the desired selling and buying prices. The advantage of this approach is that
the underlying asset is model free. Only observed stock prices are required, so
it can be performed on line. After briefly presenting the asymptotic results,
simulations and real market data are used to demonstrate the performance of
the proposed algorithm.
Keywords: Stochastic approximation; stochastic optimization; trading strategy.
1. Introduction
This work is concerned with developing a systematic numerical procedure
for trading a mean-reverting asset. The trading strategy consists of two
ingredients, buy and sell. One wishes to buy low and sell high. Nevertheless,
it is challenging to be able to correctly identify these low and high prices
246 Q. Zhang et al.
in practice. The purpose of this paper is to develop and implement an easy

while systematical procedure to determine the buying and selling prices
when the underlying asset price is subject to a mean-reverting process.
A mean-reverting model is often used in financial and energy markets to
capture the price movements that will eventually move back to an “equilib-
rium” level. Empirical studies on mean-reverting stock prices can be traced
back to the 1930s (see Ref. 10). The research was carried further in many
studies. Among them, Ref. 13 and Ref. 25 were the first to provide direct em-
pirical evidence that mean reversion occurs in U.S. stock market over long
horizon. Ref. 2 provided international evidence to support mean-reverting
stock prices in 18 countries during the period 1969 to 1996.
Mean reversion is also found in short-horizon expected returns (Ref. 7).
Other than stock prices, mean-reverting models are also used to characterize
stochastic volatility (Ref. 16) and asset price in energy market (Ref. 1).
There is a large body of literature studying trading rule in financial
markets for many years, especially on the sell side. For example, in Ref.
29 a selling rule is studied when the stock price evolves according to a se-
ries of geometric Brownian motions (GBM) coupled with a continuous-time
finite state Markov chain. An investor makes a selling decision whenever
the stock price exceeds the target price or hit the stop-loss price. The ob-
jective is to determine these threshold prices by maximizing a discounted
expected reward function, and the optimal threshold values are obtained by
solving a set of two-point boundary value problems. In Ref. 14, the trailing
stop strategy is considered for two models for stock prices: a discrete-time
random walk and continuous-time Brownian motion. A trailing stop order
maintains a stop price at a precise percentage below the market price. For
both models, they discuss the question of optimizing the percentage from
the current price to the stop. In Ref. 15, the optimal selling rule was consid-
ered for stock price under a switching GBM model and the optimal stopping
problem is solved by using a smooth-fit technique. In Ref. 24 the liquidation
problem of a large block of stocks was considered. Other than these ana-
lytical results, various numerical methods have been developed to compute
these threshold. In Ref. 28, a stochastic approximation technique was used
to obtain the optimal selling rule. Further numerical and asymptotic results
were obtained in Ref. 27. In addition, a liner programming approach was
developed in Ref. 17 and fast Fourier transformation was used in Ref. 23.
Furthermore, capital gain taxes and transaction costs in stock selling were
considered in Ref. 5, 9 and 11, among others.
On the other hand, most work on the buying side of trading is qualita-
Mean-Reverting Asset Trading 247
tive. For example, contrarian and momentum strategies were studied in Ref.
4, Ref. 8, and Ref. 18–20. Not until recently was a rigorous mathematical
analysis on the buying side provided in Ref. 30, in which they considered
buying and selling for assets governed by mean-reverting processes. The ob-
jective is to buy and sell the underlying asset sequentially to maximize the
discounted reward function when the slippage cost is taken into account.
[Slippage cost usually refers to the difference between the estimated price
and the actual price paid.] In Ref. 30, the optimal buying and selling prices
were obtained using dynamic programming approach and the associated
HJB equations for the value functions. One makes a buy decision when the
market price hits the buying price and makes a sell decision when the mar-
ket price exceeds the selling price. In order to implement the strategy in
Ref. 30, one needs to know the values of parameters in the mean-reverting
processes in order to compute the optimal threshold values. In practice, it
is difficulty to determine those values. Taking this point into consideration,
in Ref. 26, a stochastic approximation algorithm was proposed to solve the
problem for buying and selling the asset once. Instead of solving two quasi-
algebraic equations, the problem is formulated as a stochastic optimization
procedure. The algorithm is model free and uses observed stock prices only.
In this paper, we further develop the algorithm that allows buying and
selling to take place a multiple number of times.
The essential feature of our approach is the use of stochastic approx-
imation methods (see Ref. 22 and Ref. 6 for up-to-date development of
stochastic approximation algorithms). The proposed stochastic approxima-
tion algorithm allows us to deal with general model free case and use only
observed stock prices to determine the optimal buying and selling prices.
Therefore the proposed method can be easily implemented in practice.
The rest of the paper is arranged as follows. Section 2 offers a precise
formulation of the problem and the description of algorithm. Section 3
proceeds with the convergence and rates of convergence of the proposed
algorithm. To demonstrate the feasibility and efficiency of the algorithm,
numerical experiments using simulations and real market data are given
in Section 4. We demonstrate that the proposed algorithms provide sound
estimated optimal threshold values; they can be easily implemented in real
time and provide guidelines for stock trading. Finally we conclude this paper
with some further remarks in Section 5. Since the main concerns in this
paper are to develop sound numerical algorithms, and since the theoretical
development has been setup in Ref. 26, the paper is concentrated on the
numerical aspects. The detailed proofs are omitted for brevity.
May 12, 2011 11:9 WSPC - Proceedings Trim Size: 9in x 6in 11-zhang
248 Q. Zhang et al.
2. Problem Formulation
In Ref. 30, they assume that X(t) ∈ R is a mean-reverting process governed
by
dX(t) = a(b − X(t))dt + σdW (t), X(0) = x, (1)
where a > 0 is the rate of reversion, b is the equilibrium level, σ > 0 is the
volatility, and W (t) is a standard Brownian motion. Then the asset price
is given by
S(t) = exp(X(t)). (2)
In our formulation, we do not require asset price S(t) to be any specific
stochastic process or follow any specific distribution. We only assume that
asset price can be observed. Based on the observed stock price, two se-
quences of stopping times τ {bi } and τ {si } with
0 ≤ τ {b1 } ≤ τ {s1 } ≤ τ {b2 } ≤ τ {s2 } ≤ · · ·
are considered. One makes a buying decision at time τ {bi } and makes a
selling decision at time τ {si } , with i = 1, 2, .... Suppose that 0 < K < 1 is
the percentage of slippage per transaction and ρ > 0 is the discount factor.
We aim to find the optimal buying price and selling price that maximize a
suitable reward function. Thus the formulation is


 Find argmax Φ(θ) = E[J(θ)],

1 2 0
θ = (θ , θ ) ∈ (0, ∞) × (0, ∞),



∞
Problem P : X (3)

 J(θ) = [exp(−ρτ {si } )S(τ {si } )(1 − K)



 i=1
− exp(−ρτ {bi } )S(τ {bi } )(1 + K)],

where
τ {b1 } = inf{t > 0, S(t) ≤ exp(θ1 )},
τ {bi } = inf{t > τ {si−1 } , S(t) ≤ exp(θ1 )}, for i ≥ 2,
τ {si } = inf{t > τ {bi } , S(t) ≥ exp(θ2 )}, for i ≥ 1.
Note that τ {bi } and τ {si } denote the stopping times for buying and sell-
ing respectively, θ1 and θ2 denote the buying and selling threshold values
respectively, and S(t) is the stock price at time t.
The analytic solution is obtained in Ref. 30 when S(t) is governed by
(2). However, the solution depends on the values of a and b in (1), which
are difficult to be determined in practice. Our contribution is to devise a
optimization procedure that estimates the optimal threshold value θ and
only requires observed stock prices. We will use a stochastic approxima-

tion procedure (SA) to resolve the problem by constructing a sequence of
estimates of optimal threshold value θ, using
θn+1 = θn + {step size}{gradient estimate of Φ(θ)}. (4)
Let us begin with a simple noisy finite difference scheme. The only
provision is that S(t) can be observed. Associated with the iteration number
n, denote the threshold value by θn . Let us begin with an arbitrary initial
guess θ0 , we construct a sequence of estimates {θn } recursively as follows.
{b } {s }
Then we can determine stopping times τn i and τn i , buying times and
selling times as
τn{b1 } = inf{t > 0, S(t) ≤ exp(θn1 )},
τn{bi } = inf{t > τn{si−1 } , S(t) ≤ exp(θn1 )}, for i ≥ 2,
τn{si } = inf{t > τn{bi } , S(t) ≥ exp(θn2 )}, for i ≥ 1.
Define a combined process ξn that includes the random effect from S(t)
{b } {s }
and the stopping times τn i and τn i as
ξn = (S(τn{b1 } ), S(τn{s1 } ), S(τn{b2 } ), S(τn{s2 } ), ..., τn{b1 } , τn{s1 } , τn{b2 } , τn{s2 } , ...)0 .
(5)
We call ξn the sequence of collective noise. Let Φ(θ, e ξ) be the observed value
of the objective function Φ(θ) with collective noise ξ. When the threshold
± n0
values is set at θ, take random samples of size n0 with sequence {ξn,l }l=1
such that
e ξ ± ) + · · · + Φ(θ,
Φ(θ, e ξn,n ±
)
b ξ ± ) def
Φ(θ, =
n,1 0
. (6)
n
n0
We assume that
b ξ ± ) = Φ(θ) for each θ.
E Φ(θ, (7)
n
b ξ ± ) is an estimate of Φ(θ). In the simulation study,

Then for each θ, Φ(θ, n
we can use independent random samples to estimate the expected value of
b ξ ± ) converges to Φ(θ)
Φ(θn ). The law of large numbers implies that Φ(θ, n
w.p.1 as n0 → ∞. We will not assume the independence in the proof of
b ξ ± ), we will use the
convergence theorem. In lieu of using (6) with Φ(θ, n,l
b ξn ).
form Φ(θ,
To obtain the desire estimate, we construct a stochastic approxima-
tion procedure with finite difference gradient estimates. Define Yn± =
(Yn±,1 , Yn±,2 ) as
Yn±,ι (θ, ξn± ) = Φ(θ
b ± δn eι , ξn± ), for, ι = 1, 2, (8)
250 Q. Zhang et al.
where eι is the standard unit vectors with e1 = (1, 0)0 and e2 = (0, 1)0 ,
ξn± are two different collective noises taken at threshold values θ ± δn eι ,
respectively, and δn is the finite difference sequence satisfying δn → 0 as
n → ∞. We write Yn± = Yn± (θ, ξn± ). For simplicity, henceforth, we often use
ξn to represent ξn+ and ξn− if there is no confusion. The gradient estimate
at iteration n is given by
def
b n , ξn ) = (Yn+ − Yn− )/(2δn ).
DΦ(θ (9)
Then the recursive algorithm is
θn+1 = θn + εn DΦ(θ
b n , ξn ), (10)
where εn is a sequence of real numbers known as step size. A frequently

used choice of step size and finite difference sequences is εn = O(1/n) and
δn = O(1/n1/6 ). Throughout this paper, this is our default choice of step
size and finite difference sequences.
To proceed, we define
λn = (Yn+ − Yn− ) − En (Yn+ − Yn− ),
ηnι = [En Yn+,ι − Φ(θn + δn eι )] − [En Yn−,ι − Φ(θn − δn eι )], ι = 1, 2,
Φ(θn + δn eι ) − Φ(θn − δn eι )
βni i = − Φθι (θn ), ι = 1, 2,
2δn
(11)
where En denotes the conditional expectation with respect to Fn , the σ-
algebra generated by {θ1 , ξj± : j < n}, Φθι (θ) = (∂/∂θι )Φ(θ), and Φθ =
(Φθ1 (·), Φθ2 (·))0 denotes the gradient of Φ(·). In the above, ηnι and βnι for ι =
1, 2 represent the noise and bias, and λn is a martingale difference sequence.
It is reasonable to assume that after taking the conditional expectations,
the resulting function is smooth. Thus we separate the noise into two parts,
uncorrelated noise λn and correlated noise ηn . In what follows, we write
ηn = (ηn1 , ηn2 )0 and βn = (βn1 , βn2 )0 , and note that ηn = ηn (θn , ξn ). With the
above definitions, algorithm (10) becomes
λn ηn (θn , ξn )
θn+1 = θn + εn Φθ (θn ) + εn + ε n βn + ε n . (12)
2δn 2δn
3. Convergence and Rates of Convergence

This section studies the convergence and rates of convergence of the recur-
sive algorithm. We will show that θn defined in (10) is closely related to
an ordinary differential equation (ODE). The stationary points of ODE are
the optimal buying and selling prices that we are seeking.
To carry out the study of convergence, we define the following:

n−1

 X
 tn =

 εi , m(t) = max{n : tn ≤ t},
i=1 (1)


 θ0 (t) = θn for t ∈ [tn , tn+1 ), θn (t) = θ0 (t + tn ),
Nn = min{i : tn+i − tn ≥ T }, for an arbitrary T > 0.

Note that θ0 (·) is a piecewise constant process and θn (·) is its shift. With
the above definition, the interpolated process hn (·) becomes
m(tn +t)−1 m(tn +t)−1
n
X X λj
θ (t) = θn + εj Φθ (θj ) + εj
j=n j=n
2δj
m(tn +t)−1 m(tn +t)−1
(2)
X X εj
+ ε j βj + ηj (θj , ξj ).
j=n j=n
2δj
We need the following conditions:

(A1) The second derivative Φθθ (·) is continuous.
(A2) For each compact set G,
(a) supn E|Yn± I{θn ∈G} |2 < ∞.
(b) For each θ belonging to a bounded set,
n+N n −1
1 2
X
sup E 2 |En ηj (θ, ξj )| < ∞, γin | = 0, (3)
lim sup E|e
n n 0≤i<Nn
j=n
where
n+N n −1
1 X εj
ein =
γ En+i [ηj (θn+i+1 , ξj ) − ηj (θn+i , ξj )], i ≤
εn+i j=n+i
2δj
Nn − 1.
Remark 3.1. Our default choice of step size and finite difference sequences
is εn = O(1/n) and δn = O(1/n1/6 ), it follows that the sequences {εn } and
{δn } satisfy 0 < εn → 0, n εn = ∞, 0 < δn → 0, and εn /δn2 → 0 as
P
n → ∞. Moreover,
εn+i δn+i
lim sup sup < ∞, lim sup < ∞,
n 0≤i≤Nn −1 εn n δn
2
(εn+i /δn+i

)
lim sup < ∞.
n (εn /δn2 )
For simplicity, we use a mixing condition. Assume that ξn± = g0 (ζn± )
±
where g0 (·) is a real-valued function, {ζn,` } are homogeneous finite-state
252 Q. Zhang et al.
Markov chains whose transition matrices are irreducible and aperiodic.

Thus the noise is bounded since the Markov chain takes only finite val-
±
ues. Then ξn,` are φ-mixing sequences with exponential mixing rates (Ref.
3 [p.167]), i.e., $(j) = c0 $j for some c0 > 0 and some 0 < $ < 1. Us-
ing the exponential mixing rates, conditions (A2)(a) and (A2)(b) are easily
verified.
Theorem 3.2. Assume that (A1)–(A2) and that {θn } is tight in R2 . Then
θn (·) converges weakly to θ(·), the solution to the differential equation
θ̇ = Φθ (θ), (4)
which has a unique solution for each initial condition.
Corollary 3.3. Suppose that (4) has a unique stationary point θ∗ that is
globally asymptotically stable in the sense of Liapunov, and that {sn } is a
sequence of real numbers such that sn → ∞. Then θn (sn + ·) converges
weakly to θ∗ as n → ∞.
To proceed, for simplicity, take εn = 1/n and δn = δ/n1/6 . We assume

all the conditions of Theorem 3.2 holds. Define un = n1/3 (θn − θ∗ ). The
rate of convergence study aims to exploit the asymptotic properties of this
scaled sequence. We shall show that the weak limit of the interpolation of
un is a diffusion process. To proceed, let us state conditions needed in what
follows.
(A3) Assume θn → θ∗ in probability, and Φθθθ (·) exists and is continuous

in a neighborhood of θ∗ . In addition,
(a) {un } is tight;
(b) all eigenvalues of Φθθ (θ∗ ) + (1/3)I have negative real parts;
(c) for each θ,
ηn (θ, ξ) = ηn (θ∗ , ξ) + ηn,θ (θ∗ , ξ)(θ − θ∗ )

Z 1
+ [ηn,θ (θ∗ + (θn − θ∗ )s, ξ) − ηn,θ (θ∗ , ξ)]ds (θ − θ∗ );
0
(d) the sequence {ηn (θ∗ , ξn )} is stationary φ-mixing such that

E|ηn (θ∗ , ξn )|2+∆ < ∞ for some ∆ > 0 and Eηn (θ∗ , ξn ) =
0 and that the mixing measure $(·) is given by $(j) =
supA∈F n+j E (1+∆)/(2+∆) |P (A|Fn ) − P (A)|(2+∆)/(1+∆) , satis-
fying ∞ ∆/(1+∆)
P
j=1 ($(j)) < ∞.
Remark 3.4. Applying perturbed Liapunov function methods, (A3)(a)

can be verified. (see Ref. 22 [Section 10]). The existence and continuity
of Φθθθ (·) in a neighborhood of θ∗ allows us to linearize Φ(·) about θ∗ .
Conditions (A3) (c) and (d) concern about the sequence ηn (θ, ξ). Let us
examine these conditions in conjunction with (6) and using independent
±
samples and noises {ξn,l }. It is important to note that due to the use of the
stopping times τ {bi } and τ {si } , Φ(θ,
b ξ) defined in (6) may not be continuous
in θ. However, we can assume that its expectation is smooth.
e ξ) = Φ(θ)+ f0 (θ)ξ, where f0 (θ) is a bounded and
Take for instance, Φ(θ,
±
continuous function. Suppose that for a positive integer m0 , {ξn,l } are m0 -
± Pm0 ±
dependent sequences (see Ref. 3 [p.167]). For example, ξn,l = j=0 cj ζn−j,l ,
± ± 2
where {ζn,l } are martingale difference sequences satisfying E|ζn,l | < ∞.
Then they are mixing processes and the mixing measures satisfy $(j) = 0
for all j > m0 . For each θ belonging
2 to a bounded set, it is easily verified that
1 Pn0 e ±
E n0 l=1 [Φ(θ + δn eι , ξ )] < ∞. Thus E|Yn±,ι |2 < ∞. In addition, for
n,l
each j > n, En [ηj (θ, ξj )] = En {[Ej Yj+,ι (θ, ξj )−Φ(θ+δn )]−[Ej Y −,ι (θ, ξj )j −
Φ(θ − δn )]} = 0. It is easy to see that ηn (θ∗ , ξn ) is a zero mean sequence
and is φ-mixing (in fact n0 -dependent). Thus (A3)(d) is verified.
Lemma 3.5. Assume (A1)–(A3).
(a) The following inequalities hold:

∆/(1+∆)
|Eηj (θ∗ , ξj )ηk (θ∗ , ξk )| ≤ K ($(j)) ,
∆/(1+∆) (5)
E E(ηn+j (θ∗ , ξn+j )Fn ) ≤ K ($(j)) ;
Pm(t +t)−1 √
(b) The weak limit of the sequence j=nn (ηj (θ∗ , ξj ) + λ∗j )/ j is
an R2 -valued Brownian motion w(·)
e with covariance Σt,
Define a piecewise constant interpolation
u0 (t) = un , t ∈ [tn , tn+1 ), and un (t) = u0 (tn + t).
Theorem 3.6. Assume that (A3) holds. Then un (·) converges weakly to a
diffusion process u(·) that is a solution of the stochastic differential equation
δ 2 Φθ1 ,θ1 ,θ1 (θ∗ )

I 1
du = Φθθ (θ∗ ) + u+ dt + dw, (6)
3! Φθ2 ,θ2 ,θ2 (θ∗ )
e
3 2δ
e is the Brownian motion with covariance Σ1/2 (Σ1/2 )0 t = Σt given

where w(·)
by Lemma 3.5.
254 Q. Zhang et al.
Remark 3.7. To easily handle real market data, we do not take a sample
b ξ ± ) = Φ(θ,
average as in (6). Instead of that, we set n0 = 1 and use Φ(θ, e ξ± )
n n
without using sample averages. The resulting estimate is not expected to
be a smooth and the bias will be larger. Nevertheless, it does provide us a
reasonable estimate.
Define Yn± as in (8) and use algorithm (10). Since conditions concerning
b ξn ), we
the noise given in Sections 3 and 4 all refer to the function Φ(θ,
obtain the convergence and rate of convergence of the algorithm just as in
the previous sections. We summarize the above as the following proposition.
Proposition 3.1. Let n0 = 1, under the conditions of Theorem 3.2 and

Theorem 3.6, the conclusions of these theorems continue to hold.
4. Numerical Demonstration
In this section, we report our simulation results and numerical experiments.
We first use Monte Carlo simulation and compare our algorithm with the
analytic solution obtained in Ref. 30. Then we test our algorithm using real
market data.
4.1. Simulation Study

In this section, we compare the results of stochastic approximation algo-
rithm with the closed-form solution in Ref. 30. We find that the proposed al-
gorithm indeed provides good approximation results. Recall that the mean-
reverting SDE follows
dX(t) = a(b − X(t))dt + σdW (t), X(0) = x, (1)
and the stock price is given by S(t) = exp(X(t)). First, we take a = 0.8,
b = 2, σ = 0.5, x = 0, let the slippage rate K = 0.01, and the discount
rate ρ = 0.5. In this case, the analytic solution in Ref. 30 gives (θ∗1 , θ∗2 ) =
(1.331, 1.631).
We let n0 = 5000, where n0 is the number of random samples used in
each iteration; see (6). Then we use (1) to simulate the stock prices and
the recursive algorithm (10) is applied for 200 iterations. The sequence
of ξn and δn are chosen to be ξn = 1/(n + k0 ) and δn = 1/(n1/6 + k1 ),
respectively, where k0 and k1 are some positive constants. The proposed
algorithm yields the optimal
p estimation of (θ1 , θ2 ) = (1.3225, 1.6292). The
1 1 2 2 2 2
p = (θ − θ∗ ) + (θ − θ∗ ) = 0.0087, and the relative
(absoulate) error
1 2 2 2
error = error/ (θ∗ ) + (θ∗ ) = 0.0041.
Note that the analytic solutions depend on the knowledge of various

parameters: a, b, σ, K and ρ. We next vary one of the parameters at a time
and compare the approximation results with analytic results.
We first vary the values of b, the equilibrium levels of stock price. Then
we compute the threshold values (θ1 , θ2 ) associated with varying b. Intu-
itively, larger b would result in larger threshold values (θ1 , θ2 ). Our approx-
imation results confirm this. The detail results are reported in Table 1,
where θ∗1 and θ∗2 are threshold values calculated by analytic solution and θ1
and θ2 are threshold values calculated by SA method; see (10).
Table 1. (θ 1 , θ 2 ) with varying b: average error = 0.01511, average relative

error = 0.0102
b 1 1.5 2 2.5 3
θ∗1 0.331 0.831 1.331 1.831 2.331
θ∗2 0.631 1.131 1.631 2.131 2.631
θ1 0.3303 0.8211 1.3225 1.8199 2.3127
θ2 0.6504 1.1384 1.6292 2.1403 2.6407
error 0.0194 0.0124 0.0087 0.0145 0.0207
relative error 0.0272 0.0088 0.0041 0.0052 0.0059
Next, we vary a. A larger a means fast reversion rate for Xt to reach

the equilibrium level b and thus results in larger reward in short time.
Consistently, Table 2 shows that the values of (θ1 , θ2 ) increase in a.
Table 2. (θ 1 , θ 2 ) with varying a:average error = 0.0413, average

relative error = 0.0216
a 0.6 0.7 0.8 0.9 1
θ∗1 1.175 1.264 1.331 1.383 1.425
θ∗2 1.375 1.564 1.631 1.683 1.725
θ1 1.146 1.238 1.300 1.367 1.425
θ2 1.497 1.559 1.621 1.685 1.731
error 0.1254 0.0265 0.0326 0.016 0.006
relative error 0.0693 0.0132 0.0155 0.0074 0.0027
In Table 3, we vary the volatility σ. The larger the σ, the greater the
range for the stock price St = exp(Xt ). To avoid closing positions due
to normal price fluctuations, one needs to increase the values of (θ1 , θ2 ).
Consistent with this, Table 3 shows the values (θ1 , θ2 ) increase in σ.
In Table 4, we vary the discount rate ρ. A larger discount rate ρ implies
smaller return and smaller threshold values (θ1 , θ2 ). These are confirmed in
Table 4.
256 Q. Zhang et al.
Table 3. (θ 1 , θ 2 ) with varying σ: average error = 0.0339, average

σ 0.3 0.4 0.5 0.6 0.7
θ∗1 1.231 1.275 1.331 1.400 1.481
θ∗2 1.531 1.575 1.631 1.700 1.781
θ1 1.213 1.274 1.316 1.360 1.469
θ2 1.496 1.564 1.624 1.732 1.831
error 0.0394 0.0110 0.0166 0.0512 0.0514
relative error 0.0200 0.0055 0.0079 0.0232 0.0222
Table 4. (θ 1 , θ 2 ) with varying ρ: average error = 0.0638, average

ρ 0.3 0.4 0.5 0.6 0.7
θ∗1 1.681 1.456 1.331 1.206 1.081
θ∗2 1.781 1.756 1.631 1.506 1.281
θ1 1.604 1.440 1.320 1.181 1.060
θ2 1.901 1.782 1.627 1.521 1.384
error 0.1426 0.0305 0.0117 0.0292 0.1051
relative error 0.0582 0.0134 0.0056 0.0151 0.0627
Finally, we choose different slippage rates K. The results in Tables 5

suggest that θ1 is decreasing slightly in K and θ2 is almost flat. The possible
explanation is that larger slippage cost discourages stock transactions and
thus has to be compensate by smaller θ1 .
Table 5. (θ 1 , θ 2 ) with varying K: average error = 0.0656, average

K 0.001 0.005 0.01 0.015 0.02
θ∗1 1.431 1.431 1.331 1.331 1.231
θ∗2 1.631 1.531 1.631 1.531 1.631
θ1 1.413 1.351 1.316 1.264 .1252
θ2 1.585 1.619 1.624 1.633 1.634
error 0.0494 0.1189 0.0166 0.1220 0.0212
relative error 0.0228 0.0568 0.0079 0.0602 0.0104
It is clear to see from the above tables that the stochastic approxima-
tion constructed in this paper indeed provide sound estimates of optimal
threshold values (θ1 , θ2 )0 . The average error is 0.0440 and the average rel-
ative error is only 0.0220, or 2.20%.
4.2. Tests with Market Data

In this section, we study the performance of the algorithm using real market
data. The proposed algorithm are employed as follows:
Step 1. We collect stock prices during the period January 2, 2002 to

December 31, 2007;
Step 2. We divide the whole period into three sub-periods: Period
1, the first 500 trading days beginning at January 2, 2002;
Period 2, the subsequent 500 trading day following Period 1;
and Period 3, the next 500 trading days following Period 2;
Step 3. We run 600 iterations of stochastic approximation algorithm
(10) on stock prices in the Period 1 and compute the threshold
values (θ1 , θ2 );
Step 4. We use the threshold values (θ1 , θ2 ) obtained in Step 3 to
simulate trading the same stock in Period 2 and compute
the overall dollar return. Recall that we buy stocks whenever
stock price S(t) < exp(θ1 ) and sell stocks whenever stock price
S(t) > exp(θ2 );
Step 5. Again, we run 600 iterations of algorithm (10) on stock prices
in the Period 2 and compute the threshold values (θ1 , θ2 );
Step 6. We use the threshold values (θ1 , θ2 ) calculated in Step 5 to
simulate trading the same stock in Period 3 and compute the
overall dollar return.
In the above procedure, if a stock is bought in any period but the stock
price doesn’t exceed the selling price after buying, we will sell stock at
the end of the period regardless of selling price. Now we choose different
stocks to conduct above experiment. For example, Figure 1 is the graph of
historical prices of Wal-Mart Stores Inc. during the period January 2, 2002
to December 31, 2007.
Applying stochastic approximation algorithm to Period 1, the com-

puted buying price is $44.6088 and the selling price is $53.3786. Using
these threshold values in Period 2, we buy stock at $44.46 on Aug 23, 2005
and sell it at $47.12 at the end of Period 2. The dollar return is $1.7442 in
Period 2. Using stock prices during Period 2, the calculated buying price is
$42.2969 and the selling price is $47.8489. Applying these threshold values
in Period 3, we buy stock at $41.73 on Jul 14, 2006 and sell it at $47.92 on
Sep 26, 2006, and we buy it again at $42.06 on Sep 5, 2007 and finally sell
it at $48.45 on Dec 5, 2007. The total dollar return in Period 3 is $10.7784.
258 Q. Zhang et al.
Fig. 1. The Prices of Wal-Mart from Jan 2, 2002 to Dec 31, 2007
We apply the same procedure to Home Depot’s stock, see Figure 2 for
the daily stock price during the period January 2, 2002 to December 31,
2007.
Based on the prices of Home Depot in Period 1, after 600 iteration of
(10), the calculated buying price is $34.0484 and selling price is $38.9126.
Using the threshold values above, we trade Home Depot in Period 2. We
first buy it at $33.29 on Sep 2, 2004 and sell it at $39.45 on Nov 10, 2004.
We buy it again at $33.96 on May 10, 2005. However, before the ending
of Period 2, the price doesn’t raise above the selling price $38.9126 again.
Thus we have to sell it at the end of Period 2 for $32.77. The total dollar
return in Period 2 is $3.5753. Using the prices in Period 2, the computed
selling price is $32.7166 and $38.7353. Therefore, in Period 3, we buy it at
$32.61 on Dec 27, 2005 and sell it at $39.87 on Aug 9, 2007, resulting a
total profit of $6.5352.
The same procedure are also applied to other stocks. The detail trading
results are shown in Table 6. As can be seen, using the threshold values
computed by the proposed algorithm does not necessarily trigger transac-
tions in every period. However, overall speaking, the proposed algorithm in
this paper may provide trading guidelines in practice.
Fig. 2. The Prices of Home Depot from Jan 2, 2002 to Dec 31, 2007
Table 6. Trading Results

Period 2 Period 3
Stock Selling Buying Total Selling Buying Total
Price Price Profit Price Price Profit
AIG $52.15 $63.65 $13.95 $52.08 $59.94 $9.04
Du Pont $30.30 $39.50 $0.00 $36.23 $46.00 $9.07
Home Depot $34.05 $38.91 $3.58 $32.72 $38.74 $6.54
IBM $54.30 $82.13 $0.00 $71.82 $83.38 $10.98
Intel $12.63 $28.15 $0.00 $19.45 $27.00 $7.51
Microsoft $19.81 $22.24 $0.00 $21.03 $26.94 $5.90
3M $46.99 $69.21 $0.00 $69.90 $80.09 $20.06
Verizon $25.66 $31.58 $0.07 $28.83 $34.50 $7.66
Wal-Mart $44.61 $53.38 $1.74 $42.30 $47.85 $10.78
5. Further Remarks
A stochastic approximation algorithm has been developed for buying low
and selling high strategy in stock trading. Compared with the analytic so-
lution, the simulation results indicate that this algorithm provides sound
estimates for optimal buying and selling prices. One advantage of the pro-
posed method is its simple recursive form. In addition, this method only
260 Q. Zhang et al.
requires the observed stock prices. Thus this method can be easily im-
plemented in practice. As demonstrated by using real market data, the
proposed algorithm can provide useful guidelines for stock tradings.
A note of caution is in order. The approximation only works for stocks
under mean reversion. If the stock prices do not revert to an equilibrium
level, then the threshold values provided by the proposed algorithm may
make no sense in practice. Thus, before using the stochastic approxima-
tion methods, one needs to check if the mean reversion occurs in stock
prices. Finally, we note that developing a rigorous procedure to identify
mean-reverting assets is a challenge problem and maybe added to current
literatures.
References
1. C. Blanco and D. Soronow, Mean Reverting processes - Energy Price Pro-
cesses used for Derivatives Pricing and Risk Management, Commodities Now
(2001), 68-72.
2. R. Balvers, Y. Wu and E. Gilliland, Mean Reversion across National Stock
Markets and Parametric Constrain Investment Strategies, Journal of Finance,
55(2) (2000), 745-772.
3. P. Billingsley, Convergence of Probability Measures, J. Wiley, New York, 1968.
4. W.F.M.De Bondt and R. Thaler, Does the Stock Market Overreact? Journal
of Finance, 40, (1985), 793-805.
5. A. Cadenillas and S.R. Pliska, Optimal Trading of a Security when There are
Taxes and Transaction Costs, Finance and Stochastics, 3, (1999), 137-165.
6. H.F. Chen, Stochastic Approximation and Its Applications, Kluwer Academic,
Dordrecht, Netherlands, 2002.
7. J. Conrad and G. Kaul, Mean Reversion in Short-Horizon Expected Returns,
Review of Financial Studies, 2, (1989), 225-240.
8. J. Conrad and G. Kaul, An Anatomy of Trading Strategies, Review of Fi-
nancial Studies, 11, (1998), 489-519.
9. G.M. Constantinides, Capital Market Equilibrium with Personal Tax, Econo-
metrica, 51, (1983), 611-636.
10. A. Cowles and H. Jones, Some Posteriori Probability in Stock Market Action,
Econometrica , 5, (1937), 611-636.
11. R.M. Dammon and C.S. Spat, The Optimal Trading and Pricing of Securi-
ties with Asymmetric Capital Gains Tax and Transaction Costs, Review of
Financial Studies, 9, 1996, 921-952.
12. S. N. Ethier and T. G. Kurtz, Markov Processes: Characterization and Con-
vergence, J. Wiley, New York, 1986.
13. E. Fama and K. French, Permanent and Temporary Components of Stock
Prices, Journal of Political Economy, 96 (1988), 246-273.
14. P.W. Glynn and D.L. Iglehart, Trading Securities Using Trailing Stops, Man-
agement Science, 41, (1995), 1096-1106.
15. X. Guo and Q. Zhang, Optimal Selling Rules in a Regime Switching Model,
IEEE Trans. Automatic Control, 50, (2005), 1450-1455.
16. C.M. Hafner and H. Herwartz, Option Pricing under Linear Autoregressive
Dynamics, Heteroskedasticity, and Conditional Leptokurtosis, Journal of Em-
pirical Finance, 8, (2001), 1-34.
17. K. Helmes, Computing optimal Selling Rules for Stocks using Linear Pro-
gramming, Mathematics of Finance, G. Yin and Q. Zhang, (Eds), 187-198,
Contemporary Mathematics, American Mathematical Society, 2004.
18. N. Jagadeesh and S. Titman, Returns to Buying Winners and Selling Losers:
Implication for Stock Market Efficiency, Journal of Finance, 48, (1993), 65-
91.
19. N. Jagadeesh and S. Titman, Overreaction, Delayed Reaction, and Contrar-
ian Profits, Review of Financial Studies, 8, (1995), 973-993.
20. N. Jagadeesh and S. Titman, Short-Horizon Return Reversals and the Bid-
Ask Spread, Journal of Financial Intermediation, 4, (1995), 116-132.
21. H.J. Kushner, Approximation and Weak Convergence Methods for Random
Processes, with applications to Stochastic Systems Theory, MIT Press, Cam-
bridge, MA, 1984.
22. H.J. Kushner and G. Yin, Stochastic Approximation and Recursive Algo-
rithms and Applications, 2nd Ed., Springer, New York, 2003.
23. R.H. Liu, Q. Zhang, and G. Yin, Option Pricing in a Regime-Switching
Model using the Fast Fourier Transform, Journal of Applied Mathemat-
ics and Stochastic Analysis, Vol. 2006, Article ID 18109, 22 pages, 2006.
doi:10.1155/JAMSA/2006/18109
24. M. Pemy, G. Yin and Q. Zhang, Liquidation of a Large Block of Stock,
Journal of Banking and Finance, 31, (2007), 1295-1305.
25. J. Poterba and L. Summers, Mean Reversion in Stock Prices: Evidence and
Implications, Journal of Financial Economics, 22 (1988), 27-59.
26. Q.S. Song, G. Yin, and Q. Zhang, Stochastic optimization methods for
buying-low-and-selling-high strategies, Stochastic Anal. Appl., 27 (2009),
523-542.
27. G. Yin, Q. Zhang, F. Liu, R.H. Liu and Y. Chang, Stock Liquidation via
Stochastic Approximation using NASDAQ Daily and Intra-day Data, Math-
ematical Finance, 16, (2006), 217-236.
28. G. Yin, R.H. Liu and Q. Zhang, Recursive Algorithms for Stock Liquidation:
A Stochastic Optimization Approach, SIAM J. Optim, 13, (2002), 240-263.
29. Q. Zhang, Stock Trading: An Optimal Selling Rule, SIAM J. Control Optim.,
40, (2001), 64-87.
30. H. Zhang and Q. Zhang, Trading A Mean-Reverting Asset: Buy Low and Sell
High, Automatica, 44, (2008), 1511-1518.

Allanus Tsoi, David Nualart, George Yin Eds. Stochastic Analysis, Stochastic Systems, and Applications To Finance PDF

Загружено:

Сведения о документе

Оригинальное название

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Allanus Tsoi, David Nualart, George Yin Eds. Stochastic Analysis, Stochastic Systems, and Applications To Finance PDF

Загружено:

Авторское право:

Доступные форматы

Stochastic Analysis,

Stochastic Systems, and

8197.9789814355704-tp.indd 1 5/19/11 12:05 PM

8197.9789814355704-tp.indd 2 5/19/11 12:05 PM

British Library Cataloguing-in-Publication Data

STOCHASTIC ANALYSIS, STOCHASTIC SYSTEMS, AND APPLICATIONS

He Yue - Stochastic Analysis.pmd 1 5/11/2011, 3:59 PM

Contributors and Addresses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix

Part I. Stochastic Analysis and Systems

1. Multidimensional Wick-Itô Formula for Gaussian Processes . . . . . . . . 3

Part II. Finance and Stochastics

4. Real Options and Competition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

7. Jump Bond Markets Some Steps towards General Models

This volume contains 11 chapters. It is an expanded version of the papers

finding expectations of monotone functions of binary random variables by

Contributors and Addresses

• Alain Bensoussan, School of Management, University of Texas at

x Contributors and Addresses

• Shangzhen Luo, Department of Mathematics, University of North-

Multidimensional Wick-Itô Formula for Gaussian Processes

Keywords: Wick-Itô formula; Gaussian processes; Malliavin calculus.

∗ Supported by the NSF Grant DMS-0604207

4 D. Nualart and S. Ortiz-Latorre

Ri,j (s, t) = E[Xsi Xtj ],

for i, j = 1, . . . , d. For s = t, we have the covariance matrix Vt = R(t, t).

h1i[0,s] , 1j[0,t] iH = Ri,j (s, t) , 0 ≤ s, t ≤ T, 1 ≤ i, j ≤ d,

Multidimensional Wick-Itô Formula for Gaussian Processes 5

1i[0,s] = 1[0,s]×{i} (x, k) , (x, k) ∈ A.

where h1 , . . . , hm , g1 , . . . , gm ∈ H. The subspace of mth symmetric tensors

F = f (X (h1 ) , . . . , X (hn )) , (1)

with h1 , . . . , hn ∈ H, n ≥ 1, and f ∈ Cb∞ (Rn ) (f and all its partial deriva-

which is an element of L2 (Ω; H). By iteration one obtains

which is an element of L2 (Ω; H m ).

Definition 2.1. For m ≥ 1, the space Dm,2 is the completion of S with

6 D. Nualart and S. Ortiz-Latorre

The Wick product F  X (h) between a random variable F ∈ D1,2 and

F  X (h) = F X (h) − hDF, hiH .

E [F X (h)] = E [hDF, hiH ] . (2)

The Wick integral of a stochastic process u with respect to X is defined

π = {0 = t0 < t1 < · · · < tn = T }

where |π| = max0≤i≤n−1 (ti+1 − ti ) , |π|inf = min0≤i≤n−1 (ti+1 − ti ) , and

Definition 2.3. Let u = {ut , t ∈ [0, T ]} be a d-dimensional stochastic pro-

is defined as the limit in probability, if it exists, of the forward Riemann

Multidimensional Wick-Itô Formula for Gaussian Processes 7

(A3) For all j, k ∈ {1, . . . , d}

where ∆i X j = Xtji+1 − Xtji , and π runs over all partitions of [0, T ]

Our purpose is to derive a change-of-variable formula for the process

(A4) For every multi-index α = (α1 , ..., αd ) ∈ Nd with |α| := α1 + · · · +

exist, are continuous, and satisfy

for all t ∈ [0, T ] , x ∈ Rd , where CT > 0 and cT are such that

(see Lemma 4.5).

8 D. Nualart and S. Ortiz-Latorre

exist and the following Wick-Itô formula holds:

Proof. Using the Taylor expansion of f up to fourth order in two consec-

Multidimensional Wick-Itô Formula for Gaussian Processes 9

where δi = (ti , ti+1 ]. Taking into account that

Using the definition of h·, ·iH and adding and subtracting 21 E ∆i X j ∆i X k

10 D. Nualart and S. Ortiz-Latorre

Using Assumption (A1) it is easy to show the almost sure convergence

and the result follows.

Multidimensional Wick-Itô Formula for Gaussian Processes 11

Remark 3.1. We can also consider a function f (t, x) depending on time

In order to prove Propositions 5.1, 5.2 and 5.3 we need to introduce

which is an element of L2 (Ω; Hm ).

The Wick product F X (h) between a random variable F ∈ D1,2 and

F X (h) = F X (h) − hDF, hiH .

×h(11[0,ti ] )u1 · · · (1d[0,ti ] )ud (11[0,ti ] )v1 · · · (1d[0,ti ] )vd ,

(11[0,ti ] )u1 · · · (1d[0,ti ] )ud (11[0,ti ] )v1 · · · (1d[0,ti ] )vd

E[∂ u (∂j21 ,k1 f (Xti ))∂ v (∂j22 ,k2 f (Xti2 )] ≤ aT < ∞.